Optimizing for cost and intelligence

📝 摘要 #

本文是 Anthropic 官方关于 Claude Platform 上平衡成本与智能的综合指南，系统梳理了从原型到生产环境中控制 LLM 调用成本的两大类杠杆——"免费收益"（不牺牲质量的降本手段）和"质量权衡"（用智能换成本的选择），并附以大量内部基准测试的量化数据。

核心概念： LLM 生产部署的成本优化并非一味选便宜模型，而是理解每个成本杠杆对输出质量的影响——有些杠杆（如 prompt caching）降本不降质，有些（如 effort、模型选择）则需要在成本-质量曲线上找到最佳工作点。

关键技术栈 / 关键要点：

1. Prompt Caching 是最大杠杆 — 在 agent loop 中，任务成本随轮次平方增长，caching 将重复前缀的计费降至输入价的 1/10，实测将 agent loop 成本降低 2.5–3.7 倍，小 triage agent 降本 83–88%
2. Effort 参数是单模型最直接的旋钮 — low 在知识工作基准上仅损失 1–3 分即可降本 1/3–1/2；medium 通常以默认 70–85% 的成本匹配默认精度；但长编码任务（SWE-bench Pro）是真实权衡，low 损失约 8 分
3. 失败重跑策略 — 低 effort 跑全部 + 高 effort 重跑失败项，在 SWE-bench Pro 上以半价维持相同通过率
4. Advisor 策略（小模型主循环 + 大模型顾问） — 适合串行工作流中少数关键决策点，成本取决于 consult rate；低 effort executor 容易"不问顾问"，需 prompt 引导
5. Orchestrator 策略（大模型主循环 + 小模型工人） — 仅在两种场景省钱：(a) 常规任务的长尾成本保险，(b) 输入超过单上下文窗口的大规模分区扫描；单上下文内的工作不如直接降 effort
6. Token 卫生与 Prompt 审计 — 动态过滤、图像缩放、tool search、context editing、compaction 等手段组合可进一步降本；针对当前模型的 prompt 审计可免费移除过时指令，Opus 4.8→5 迁移实测降本 14% 同时提准
7. Task Budget 与 max_tokens 的区别 — budget 是模型可见的倒计时，引导自我调节；max_tokens 是不可见的安全帽，过低会截断有效输出而不省钱

关键结论 / 实践建议： 优化顺序应为：(1) 开启 prompt caching + token 卫生 → (2) 在当前模型上扫 effort → (3) 如 effort 无法闭合差距再考虑多模型架构。任何多模型配置都必须与单模型的全 effort 曲线比较，因为多次实测中，降低 effort 比引入第二个模型更便宜。

When a workload moves from prototype to production, cost becomes a first-class design constraint. The most capable model can be too expensive at scale, and the least expensive model can fall short on quality. Managing cost well means understanding how each cost lever affects output quality, because some levers trade against quality and some don't. The Claude Platform gives you direct control over that tradeoff. You choose the model, the effort level, and the architecture for each request, which lets you place a workload almost anywhere on the cost-to-intelligence frontier.

The levers come in two kinds:

• Free wins cut spend without touching quality: prompt caching, token hygiene, a prompt audit against the model you are running, batch processing at 50% off for work that can wait up to 24 hours, and workspace spend limits as the backstop.
• Tradeoffs exchange cost for intelligence: model choice, effort, output caps and task budgets, and multi-model architectures.

Each lever comes with measured results and the rule for when it pays. In Anthropic's measurements, prompt caching was the largest lever by a wide margin: it cut agent-loop cost by a factor of 2.5 to 3.7 on this guide's benchmarks and cut a small triage agent's bill by 83%, or 88% with input trimming added. The multi-model levers are narrower; a second model paid off in two shapes, an advisor and an orchestrator.

Start here #

Match your situation to a row.

These results are Anthropic-internal (Benchmarks referenced) and directional, not guarantees, so measure on your own workload with the four-step method.

Cut spend without losing quality #

Prompt caching, token hygiene, batch processing, and a prompt audit against your current model all lower what you pay without lowering output quality. Two caveats apply: batch processing trades latency for its discount, and context editing, a token-hygiene lever, cost more than it saved in the run measured in this section.

Cache repeated context #

Turn on prompt caching before any other lever, because every turn of an agentic task resends the entire growing conversation: system prompt, tool definitions, and every prior turn. A 40-turn task sends its first turn 40 times, so task cost grows with roughly the square of turn count. Caching does not stop the resending, but each resend costs about a tenth as much and processes faster: the prefix is billed at the cache-read rate, a tenth of the input price, and each turn pays the 1.25x cache-write rate only for what is new.

Across Anthropic's measured runs, cache reads are routinely the largest single component of task cost, making caching worth more than most model-choice decisions. Anthropic priced WideSearch ¹ and DeepResearch Bench II ⁷ runs with and without caching:

The cache's default lifetime is 5 minutes and an agent loop's turns are seconds apart, so the discount applies to most tokens on every turn; the charted runs achieved 81% to 90% hit rates. The saving varies with episode depth, because shorter loops re-read less, but caching stayed the largest single lever on every model and benchmark measured.

If your loop waits on humans between turns, use the 1-hour cache duration. It costs more to write (2x the input price instead of 1.25x) but pays for itself on the first prevented miss, because a miss resends the whole prefix at full price and writes it again.

Setup takes little work. Automatic caching places breakpoints for you; otherwise, the Claude API skill that ships with Claude Code can add caching to an existing integration from one prompt. The following excerpt shows the skill adding it to the harness that produced these measurements:

$ claude
> add prompt caching to this integration

Done. Prompt caching is now wired into the harness. Two changes:

- build_system() now adds cache_control: {"type": "ephemeral"} to the
  system block. Since the API renders tools -> system -> messages, this
  one breakpoint caches the entire static prefix -- the full tool array
  plus the system prompt.
- apply_moving_cache_breakpoints() strips any stale markers, then marks
  the last content block of the two most recent user turns. The older
  marker is the read point matching the prefix the previous request
  cached; the newer one extends the cache for the next request.

That's 3 breakpoints total, under the limit of 4.
...

Those breakpoint placements follow the standard pattern in Explicit cache breakpoints.

Three settings can break your cache during a task. Changing effort between requests invalidates the cached prefix, so change it only where you would re-cache anyway, such as at a compaction boundary. Changing a task budget partway through does the same, so set it once, on the first request. Every context editing pass invalidates the prefix from the point it clears and the next request pays to re-cache everything after it, so clear in a few large batches rather than many small ones. Make all three changes at natural breaks, then confirm cache reads have not dropped; if they have, cache diagnostics shows where the prefix diverged.

Trim input and context tokens #

Most agent requests carry tokens that never influence the answer. Trimming them costs nothing in output quality, although not every lever here saved money when measured. Two places to look:

• Input trimming. Dynamic filtering in the web fetch tool keeps boilerplate out of fetched pages, image resizing right-sizes vision inputs, and tool search with deferred loading loads tool definitions only when needed. Programmatic tool calling lets Claude run several tool calls from code so only the filtered result enters the context; its documentation reports 24% fewer input tokens on agentic search benchmarks, with a higher score. Manage tool context compares tool search, programmatic tool calling, prompt caching, and context editing.
• Context lifecycle. Context editing clears stale tool results, and automatic compaction with its threshold stops long loops from carrying their whole history forward.

The levers interact with the cache and each other, so judge them by net effect, and use cache diagnostics to confirm your cached prefix survives each change. Anthropic turned the levers on one at a time for an issue-triage agent working through 20 real bug reports with screenshots from a public repository (and, for the second panel, a longer variant of the same job):

Caching did almost all the work, and trimming took the total to 88%. Each bar is one run, so differences of $0.10 are noise; those shown here are not. Compaction needs a session long enough to trigger it: the 20-issue run never reached the 50,000-token floor once its inputs were trimmed, but on the second panel's longer variant it fired once and cut the bill a further 38%.

Context editing is the one lever here that is not free. Every clearing pass rewrites the cached conversation, which works against prompt caching; in this run, context editing cost more than it saved. Use it to make room in the context window, and clear in a few large batches.

Batch work that can wait #

The Batch API takes 50% off every token of a request, including cached ones, in exchange for results arriving any time within 24 hours. Route every request no one is waiting on through a batch, and keep the interactive path for the rest. Batching is the second-largest free lever after caching for unattended agent work: evaluation runs, backfills, and scheduled jobs such as a recurring run of the issue-triage agent from the token-trimming measurement. It combines with everything on this page except interactivity, but is not available for Claude Managed Agents sessions, which are interactive by design (see Claude Managed Agents pricing).

Audit prompts against the current model #

Each model generation responds to prompts differently, so a prompt accumulates text written for a model you no longer use. The usual case is over-specific instruction added to compensate for an older model: "verify twice," "be maximally thorough," a mandatory step-by-step procedure, or a hand-rolled reasoning scratchpad. A newer model follows these to the letter, producing extra tool rounds and extra writing, so the bill goes up with no gain in accuracy. Auditing prompts against the model you run now, and again whenever you change models, is a free win.

The audit is one command. The Claude API skill that ships with Claude Code has a prompt-audit command that reads a project's prompts and request code and reports what was written for a different model. This shortened excerpt shows it run against a support-desk prompt and request code containing those patterns:

$ claude
> /claude-api prompt-audit

10 findings. The three that matter most:

1. budget_tokens returns a 400 on Opus 5 (support_agent.py:10). This code
   does not run against its own target model. Everything else here is
   quality; this one is an outage.
2. The prompt contains two live contradictions. Line 15 says confirm before
   every refund, line 17 says process every eligible refund immediately.
   Line 19 asks for a complete recap *and* a three-sentence maximum.
3. The reasoning scaffold and the 6-step script fight the model rather than
   steer it. <scratchpad> + "reason step by step" is now a request
   parameter, not prose; the mandatory 6-step procedure plus "investigate
   fully even when the ticket looks simple" forces four tool calls on a
   "where's my package" ticket.
...
-After any refund or escalation, verify twice before submitting: re-fetch
-the order, re-check every figure in your reply against the fresh lookup,
-and review the reply a second time for errors.
+Before submitting a refund or an escalation, re-fetch the order and confirm
+every figure in your reply matches the fresh lookup.

The command then proposes its edits as a diff (one hunk shown) and lists what it deliberately left alone: the refund window, the tone requirement, and the quality bar. You review a patch, not a rewrite.

The effect is measurable. On a support-desk evaluation ¹⁴, prompts written for Claude Opus 4.8 cost 36% more per ticket on Claude Opus 5 for no change in accuracy. Running the audit over the same prompts made Opus 5 both cheaper than the unaudited version (by 14%) and more accurate (97% of tickets, up from 92%, a gain outside the noise). On the Claude Sonnet 4.6 to Claude Sonnet 5 migration, the audit took 14% off at the same accuracy:

The two kinds of stale text have different costs. Instructions the new model follows too literally cost money: removing "verify twice" cut Opus 5's cost per ticket by a third, and removing "be maximally thorough" almost as much. Text that no longer fits the model costs accuracy instead: a retired thinking setting, contradictory rules, and a hand-rolled scratchpad that conflicts with the model's own thinking each restored 7 to 11 points on Opus 5 when removed:

The same patterns appear in tool descriptions and skills, and are worth removing there too.

Trade cost against intelligence #

These levers set where a single model sits between cost and intelligence: model choice, effort, re-running failures at a higher setting, and the budgets and caps it works within. Start with an effort sweep on your current model (Tune effort). From lowest to highest cost and capability, the current models are Claude Haiku 4.5, Claude Sonnet 5, Claude Opus 5, and Claude Fable 5 (the frontier model); Models overview has the full lineup and prices.

Compare models on cost per task #

Price lists are written per token, and per token the frontier model looks expensive: Claude Fable 5's per-token price is several times Claude Sonnet 5's. You pay for completed tasks, though, so compare models on cost per completed task. A more capable model finishes a task with less work: fewer turns, less searching, less re-reading of its own context, and less backtracking. The per-token premium is routinely overwhelmed by doing less of everything.

Anthropic measured this directly on DeepResearch Bench II ⁷, a research-report benchmark hard enough to separate the models:

The frontier model at low effort was more accurate and about 10% cheaper per task than the mid-tier model, despite the per-token gap. It does not always win, though. On this page's SWE-bench Pro ³ subset, which both models largely saturate and whose scores are not comparable to the public leaderboard, Claude Opus 5 alone matched Claude Fable 5 alone (91.7% compared with 91.3%, inside run-to-run noise) at about 60% of its cost. On harder work, such as the DeepResearch Bench II tasks, Fable's advantage reappears.

For most agent workloads, start with Claude Opus 5: per token it costs half what Fable 5 does and 2.5 times what Sonnet 5 does, and on that coding subset it matched Fable's accuracy. At the other end, Claude Haiku 4.5 answered GPQA Diamond ⁹ questions at about a tenth of Opus 5's cost per question, with 63% accuracy compared with 92% for Opus, and fell much further behind on long coding tasks. It fits high-volume work with checkable outputs, not long agentic loops.

The ranking flips by workload, and no price list tells you which way. Price every candidate in cost per completed task on your own traffic, including Claude Opus 5 and the frontier model at reduced effort.

Price the tail of your workload, not the median: compare models on the hardest tenth of your tasks, not the typical one. On the typical task every model looks similar and the cheapest looks best, but the bill is decided by the tasks the cheaper model fails, because a failed task still bills its tokens, then the retry, then whatever the failure costs downstream. The tail is also where the money goes even when nothing fails. On a 20-problem WideSearch ¹ run, two problems carried 43% of the spend:

The multi-model strategies exist to spend frontier intelligence on that tail without paying frontier rates for the rest.

Tune effort #

Effort is the most direct way to tune a model to your task. The effort parameter governs how much thinking, tool calling, and self-verification the model does, and the default (high) suits demanding tasks. Cost scales with all that activity; accuracy scales only with the part your task needs. Below the model's ceiling, the highest effort levels pay for depth the task never uses.

On the research and knowledge-work benchmarks (WideSearch ¹, DeepWideSearch ⁶, BrowseComp ⁴, and GDPval ², all with Claude Fable 5), the curve of accuracy against cost is nearly flat: low gave up 1 to 3 points for a third to a half off the cost per task, medium matched the default's accuracy at 70% to 85% of its cost, and the default bought nothing measurable over medium on any of the four. On DeepWideSearch, low also matched an orchestrator with a Claude Sonnet 5 worker at 20% lower cost: lowering effort beat an architecture change.

Lower settings are also faster, which matters when latency is the constraint. In these runs, low took 4.5 minutes per problem on DeepWideSearch, compared with 7.9 minutes at the default. On the corpus benchmark, whose input does not fit in any single context window, Fable 5 took 7.9, 9.1, and 11.4 hours per episode at low, medium, and the default.

Long-horizon coding is the other shape. On SWE-bench Pro ³, Claude Opus 5 gave up about 2 points at medium for half the cost and about 8 points at low for a quarter of it: a real tradeoff, which re-running failures at higher effort turns back into a saving. This chart plots accuracy against cost for the research and knowledge-work benchmarks and for SWE-bench Pro:

Two consequences follow. First, draw this curve for your own workload before you add a second model: in these internal measurements, a multi-model configuration that looked cheaper than the default single model cost more than that same model at lower effort. Second, this curve is the single-model baseline any multi-model strategy must beat, so step 2 of measuring on your own workload baselines across effort levels.

Lower effort does cost accuracy on workloads that reach the model's ceiling, where accuracy genuinely scales with reasoning depth. On DeepResearch Bench II ⁷, where each report rewards deep reasoning per subtopic, every effort step bought about 2.4 points of rubric score; there is no free cost cut on that curve:

The task description alone does not reveal which kind of workload you have, so sweep two or three effort levels on a sample of your own traffic and read the answer off the curve. Test each level in a separate session: changing effort mid-session invalidates the cache (see Cache repeated context) and distorts the comparison. For parameter details, see Effort.

Re-run failures at higher effort #

When a task's outcome is checkable, the cheapest policy on the effort curve is not a fixed setting: run every task at a low setting and re-run only the failures at a higher one.

Anthropic computed this policy task by task from the effort runs on the SWE-bench Pro ³ subset in Tune effort. With Claude Opus 5 at low, 16% of tasks failed; with those re-run at the default, about 93% passed for about $0.70 each, against 91.7% for $1.39 running everything at the default: the same pass rate for half the cost, counting the failed cheap attempts. Starting at medium instead solved about 94% for about $0.95. Most of the small lift is the second attempt (re-running the default's own failures at the default scores about the same, for more money), so use this policy for the saving, not the lift:

Two conditions apply. First, you need a failure signal (here, the benchmark's own tests); a checker that passes bad work lets those failures through. Second, every first-pass failure takes two runs' worth of wall-clock time, so the saving is paid for in latency on the failures.

Set budgets and output caps #

Most agentic task runs are cheap, but a minority spend many times the median cost on searching, re-verifying, and over-testing. A task budget targets that tail. The model sees a live token countdown for the whole task and self-regulates, trimming low-value searches, skipping redundant verification, and wrapping up instead of spiraling.

Anthropic measured pass rate and cost per task on SWE-bench Pro ³ with Claude Fable 5 as the budget tightened:

A generous budget gave up about 2.7 points of pass rate for an 18% cost saving, and the tightest allowed budget gave up 4.4 points for a 47% saving. Budgets bought efficiency here, not accuracy.

Three controls do three different jobs. A task budget saves money, because the model sees it. max_tokens is a safety cap that saves nothing. On Claude Managed Agents, a session budget is the hard dollar stop behind both. Set all three: a task budget, a high max_tokens, and a session cap for the run you never want on a bill, with a workspace spend limit as the final backstop.

• Task budgets are in beta (beta header task-budgets-2026-03-13) on Claude Opus 5, Claude Fable 5, Claude Opus 4.8, and Claude Opus 4.7, but not Claude Sonnet 5; check the support table first. Start near your loop's 90th-percentile token usage, then tighten (Choosing a budget shows how to collect that distribution). Budgets below the current 20,000-token floor are rejected, and very tight budgets can produce refusal-like behavior. Set the budget once, on the first request, because a mid-task change invalidates the cache. The budget is advisory, steering the model rather than stopping it, so verify adherence on your workload.
• max_tokens caps a single response, invisibly to the model, so lowering it does not make the model economize. The turns that needed the room are discarded and still billed. On an internal repository-task benchmark ¹², a 16,384-token cap ended 15% of Claude Opus 5's attempts and a third of Claude Fable 5's, none of them solved. Capped runs spent less per attempt but bought proportionally fewer solves, so cost per solved task was the same as at 64,000. At that setting nothing was cut off, and Fable solved 54.6% of tasks instead of 33.3% (on a separate cut of the SWE-bench Pro ³ subset, described in reference 12, 92% instead of 90%). Retrying capped attempts only adds cost: at the same cap they never succeeded, and at a higher one you also pay for the wasted attempt. Set max_tokens to 64,000 for agentic work (128,000, the maximum, at xhigh or max effort), stream responses that large, treat stop_reason: max_tokens as a failure, and save money with effort and task budgets, which the model can see.
• Session budgets on Claude Managed Agents are the hard stop. A session budget is a dollar cap on one session at list rates for tokens, searches, and session time. At the cap, the session pauses with stop_reason: budget_reached; raising the budget resumes it. It is platform-enforced, works on any model with a list price (including Claude Sonnet 5), and combines with the advisory task budget. Deployments apply the same field to every run.

The first of two max_tokens charts plots cost per attempt and per solved task at each cap:

The second plots per-turn output length against the caps:

Combine models #

Multi-model architectures fit workloads whose task complexity varies enough that different steps are best served by different models. When your traffic mixes routine work that a smaller model handles reliably with harder steps that need frontier capability, splitting the work keeps frontier intelligence where it matters while most tokens bill at smaller-model rates. When a workload lacks that mix, because its difficulty is uniform or it is one dependent chain, a single well-tuned model is usually the better choice. Each strategy section gives the rule for telling the two cases apart.

Two strategies cover most workloads, and they differ in which model holds the main loop:

Advisor strategy: escalate hard decisions #

In the advisor strategy, a lower-cost executor model runs the agent loop and performs most turns. When it hits a decision that needs deeper judgment, such as choosing an approach or recovering from a failure, it calls a higher-intelligence advisor model for strategic guidance, then continues. Most tokens are billed at executor rates, and only the occasional consultations at advisor rates.

To use it, add the advisor tool to your request. This beta feature runs the whole strategy server-side in one /v1/messages request: the executor emits a tool call, Anthropic runs the advisor inference, and the executor continues with the advice; you write no orchestration code. On Claude Managed Agents, give the session an advisor by adding an advisor entry to the agent's multiagent roster; the session's primary thread consults it the same way. Claude Code supports it too; see escalating hard decisions with the advisor tool.

What sets the payoff. The advisor sees the task only through the executor's calls, so two things decide how much it helps.

The first is the gap between the models. The advisor can only hand over capability the executor lacks: on GPQA Diamond ⁹ a Claude Haiku 4.5 executor gained a great deal from a Claude Opus 5 advisor, a Claude Sonnet 5 executor gained a few points, and a frontier executor almost nothing.

The second, and the fragile one, is whether the executor actually asks (the consult rate). An executor at low effort can stop noticing it is stuck: a pairing that consults on most tasks at the default effort can fall to consulting on almost none when effort is lowered, and then scores below the executor alone. The rate also varies by task: on DeepSWE ¹⁰ a low-effort Sonnet 5 executor kept asking and gained 23 points; on SWE-bench Pro ³ the same executor stopped. When the executor does ask, it gets most of the way there. Across the pairings in the following chart, the advisor closed 60% to 90% of the gap to the stronger model while that model was paid for only on the consultations, which makes the cost cases possible:

The consult rate responds to prompting. With only the tool's built-in description, executors under-call, especially on coding work, so the advisor tool documentation gives a system prompt that asks for one call before substantive work and one before finishing, about two to three calls per task. The coding pairing measured next ran at that cadence, about two consultations on every task. That page also covers nudging an under-calling executor and capping calls client-side to bound cost. So watch the consult rate: prompt for it, measure it, and restore the executor's effort if it collapses.

When it pays on cost. An advisor saves money when a few short consultations, billed at the advisor's rate, replace running the advisor's model for the whole task. That works best when the advisor's model is priced well above the executor's, so the most cost-effective configuration is a frontier advisor over a mid-tier executor. A pairing can hold its own even at the top of the range, because advice also saves executor tokens: an executor told the right approach explores fewer dead ends, which can cover the consultations.

On an internal agentic-coding benchmark ¹¹, run with a plain API agent, a Claude Opus 5 executor with a Claude Fable 5 advisor was the most accurate configuration measured: 85.7% of attempts solved for $8.40 per attempt. It sits above the line through each model's own effort settings, but only a point or two over the best of them (Opus alone at the default setting, 84.4% for $8.50, and Fable alone at medium, 83.4% for $8.20), which a single run does not separate from noise. Fable alone at medium effort reaches about the same accuracy as the pairing (83.4% compared with 85.7%) for about the same money ($8.20 compared with $8.40 per attempt):

An earlier measurement through Claude Code's advisor mode produced the same ordering. Read this result as a shape to test on your workload, not a saving: at the top of the range the advisor buys a little accuracy at the frontier price, and the cost case belongs to pairings with a wider capability gap, such as the following chart-reading case. The latency cost is the consultations themselves: about two extra frontier-model calls per task on this benchmark, each on the task's critical path.

When it pays instead of raising effort. Where the capability gap is wider and a workload's accuracy responds to effort, a low-effort executor that consults an advisor can be a cheaper step up than raising the executor's own effort, because the advisor is paid for only on tasks that need it.

On Chartography ¹³, a public chart-reading benchmark run on Claude Managed Agents with its advisor, a Claude Opus 5 executor at low effort with a Claude Fable 5 advisor scored 67.5 for $0.60 per task. That is above the line through either model's own effort settings (Opus alone went from 49 to 75 between low and medium for $0.38 to $0.94), although the executor's own medium and default settings still hold the top scores, at 1.6 and 3.3 times the price:

The low-effort executor consulted the advisor on 86% of tasks, the condition the SWE-bench Pro pairing failed to meet. Measure the consult rate in your agent loop before relying on this configuration.

Whatever the pairing, first price the advisor's model alone at low effort; that is the baseline to beat. Recheck at every model release, because releases move both the capability gap and the price ratio.

When it fits. The advisor strategy suits workloads where turns are mostly mechanical but an excellent plan matters: coding agents, computer use, and multistep research pipelines. It fits poorly when every turn genuinely needs frontier capability, when there is nothing to plan (single-turn Q&A), or when your executor is already close to the advisor's capability.

Orchestrator strategy: delegate bulk work #

In the orchestrator strategy, the frontier model holds the loop. It decomposes the task, dispatches subtasks to lower-cost worker models, and merges their results. The orchestrator's own transcript stays short because workers absorb the token-heavy exploration, so most tokens are billed at worker rates while the plan and synthesis still come from the frontier model.

To build one, use multiagent orchestration in Claude Managed Agents: configure a coordinator agent (the orchestrator) and a roster of worker agents, each with its own model. For a complete working example with a Claude Fable 5 coordinator and Claude Sonnet 5 workers, see the Claude Cookbook recipe Coordinator pattern: big models for planning, small models for execution.

This pattern saves wall-clock time when workers can run in parallel: on the corpus benchmark ⁸ later in this section, an episode took 1.9 hours with the coordinator compared with 11.4 hours solo. It saved money in only two measured situations. On work a single model could handle alone, the same model at lower effort was cheaper every time.

Case 1: insurance against the cost tail on routine work. A frontier model running alone occasionally spirals on a routine problem it would normally solve. Because you cannot tell in advance which those will be, a few such runs dominate the bill. A coordinator that hands routine work to a lower-cost worker caps that tail, because any spiraling now happens at worker rates.

Anthropic measured this on a deliberately easy slice of BrowseComp ⁴ (10 problems the solo model reliably solves; 50 delegated and 70 solo runs). A Claude Fable 5 coordinator with one Claude Sonnet 5 worker cost a little under half as much as Fable alone on average and about a third as much at the 90th percentile ($12 compared with $33), and the solo model's single most expensive run, at $84, was also wrong:

Delegation paid on the routine, normally solvable share of the work, the opposite of the intuition that workers are for hard problems. On the full, harder BrowseComp set, the economics reversed. If your traffic has a long cost tail on routine tasks, this is the orchestrator case to measure first.

Case 2: work larger than one context window. A solo model must work through an input that large serially, one context window at a time, paying to re-read its own state on every pass. Workers each read their own partition, in parallel and at worker rates. Reading-heavy work that still fits in one context window is a model-choice problem, not a delegation problem: on reading cost alone, the orchestrator comes out ahead only when no single context can hold the work.

Anthropic built a benchmark for this case ⁸: a 21.6-million-token corpus of 14 public Python packages with 130 planted defects, too large for any context window. Lowering effort cannot help, because the bill is the corpus read itself: Claude Fable 5 solo cost $720 to $764 per episode at every effort setting, and only its accuracy moved. The coordinator configuration cost 55% less than any of those settings and scored 3 to 7 points below Fable at medium or the default, while beating a Claude Sonnet 5 solo baseline outright:

The token accounting shows why. The coordinator configuration read more than the solo model (32 million input tokens compared with 14.5 million) and still cost less, because partitioned reading at worker rates is cheaper than repeated re-reading at frontier rates. Fable 5 at default effort still holds peak accuracy, at 2.3 times the coordinator configuration's cost, so delegation here buys most of the accuracy, not all of it.

When delegation doesn't pay. An orchestrator buys something only when there is bulk to hand off: many independent pieces, ideally too many for one context window. When the work is one dependent chain, or fits in a single context, the orchestrator pays for a plan, a handoff, and a merge that a single model gets for free. In every such case measured, the coordinator's model alone at lower effort came out ahead.

BrowseComp ⁴ shows the boundary inside one benchmark. Delegation paid on the routine slice and lost on the full, harder set, where the frontier model alone reached the coordinator configuration's accuracy at 22% to 30% lower cost. Independent external work reports the same pattern ⁵. If the work is one chain, fits in one context without a long cost tail, or a single model at lower effort already meets your bar, don't build an orchestrator.

Choose between the strategies #

Most cases come down to one question: does the work split into independent pieces, or is it one answer reached through a chain of dependent steps? The strategy table maps the two answers to the two strategies.

If you are unsure, don't build anything yet:

1. Sweep effort on your current model first. It is the cheapest experiment on this page, and most workloads end there.
2. If the sweep shows a gap, price the stronger model alone at low effort. That is the number an advisor pairing has to beat, and the pairings on this page that beat it were the ones whose executor actually consulted.

The multi-model results on this page were judged against the same model at lower effort and against the next model down running alone. That is the comparison to run on your own workload, and why the first step is an effort sweep.

When you do add an advisor, it is a tool definition rather than a rearchitecture.

Measure on your own workload #

The numbers on this page are from July and August 2026, at the list prices of the time, and will drift as models and prices change. Your escalation rate, how cleanly tasks split, and transcript length move them too. The method stays the same:

1. Pull a few tasks from production logs, weighted like real traffic, and write outcome checks for each: tests pass, ticket closed, row count correct. Record cost per task beside the score: price the four token counts in each response's usage at their own rates, summed across the task's requests (the Usage and Cost API reports the aggregate).
2. Baseline the model tiers across effort levels, not only the default, and plot score against spend. A multi-model configuration must beat the single model's whole curve.
3. If the curve shows a gap effort can't close, add the multi-model strategy that fits and re-run the suite.
4. Run the winner in shadow on a traffic slice before cutover, then keep the suite running.

The following example computes one request's step 1 cost at Claude Opus 5's list prices:

# Per-million-token prices from the pricing page; change these two for another model.
INPUT_PER_MTOK = 5.00  # Claude Opus 5
OUTPUT_PER_MTOK = 25.00

client = anthropic.Anthropic()
response = client.messages.create(
    model="claude-opus-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello, Claude"}],
)
usage = response.usage
cost = (
    usage.input_tokens * INPUT_PER_MTOK
    # Cache writes bill at 1.25x the input price (5-minute cache); cache reads at 0.1x.
    + (usage.cache_creation_input_tokens or 0) * INPUT_PER_MTOK * 1.25
    + (usage.cache_read_input_tokens or 0) * INPUT_PER_MTOK * 0.10
    + usage.output_tokens * OUTPUT_PER_MTOK
) / 1_000_000
print(f"Request cost: ${cost:.6f}")

In agent loops the cache-read term is usually the largest of the four; if not, check that caching is engaged. When the advisor tool or compaction is enabled, some tokens are reported only in usage.iterations and not in the top-level totals, so sum over usage.iterations instead, pricing advisor_message entries at the advisor model's rates.

The following table lists the levers in the order to try them:

Benchmarks referenced #

All measurements are Anthropic-internal runs of these benchmarks. Unless noted, costs are USD at August 2026 list prices; Claude Sonnet 5 figures use $2 and $10 per million input and output tokens. Charts labeled "notional USD" price each request's token counts at those rates rather than reporting invoices.

1. WideSearch: Wong et al., "WideSearch: Benchmarking Agentic Broad Info-Seeking," arXiv.07999, 2025. Broad web-research tasks graded on a many-row table's completeness and accuracy; 200 problems, 3 runs per configuration. The caching and effort charts come from separate runs, so per-problem costs differ slightly. The cost-concentration chart is a separate 20-problem run, 3 runs per problem, costed from per-request billing records.
2. GDPval: OpenAI, "GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks," 2025. Knowledge-work deliverables graded against task rubrics; a 210-task run of the released gold set, one attempt per task. A Claude model grades, so absolute scores may differ from published results.
3. SWE-bench Pro: Scale AI, "SWE-Bench Pro: Can AI Agents Solve Long-Horizon Software Engineering Tasks?", 2025. A 482-problem subset selected for compatibility with Anthropic's evaluation harness; scores are not comparable to the public leaderboard. Claude Opus 5 at the default effort averages two runs; reduced-effort settings are single runs. Escalation figures come task by task from those runs: low first, then the default on its failures, solved 92.5% to 93.6% across run pairings for about $0.70; medium first, 93.8% to 94.2% for about $0.95; the default re-run on its own failures, 94.0% for $1.58; everything at the default, 90.9% to 92.5% for $1.39. The Claude Sonnet 5 executor pairings on the advisor chart come from the same August 2026 series on this subset: the Sonnet-plus-Opus pairing was run twice (a run and an exact replication) and the low-effort pairing once; the task-budget figures are one run per budget on the same subset. The Claude Fable 5 figure in Compare models is a single July 2026 run, also the task-budget chart's unbudgeted baseline; every budgeted run completed all 482 problems without harness errors.
4. BrowseComp: Wei et al., "BrowseComp: A Simple Yet Challenging Benchmark for Browsing Agents," OpenAI, 2025. Effort figures use a 500-problem cut, one to three runs per setting. The cost-insurance chart uses 10 reliably solved problems from a 26-problem slice, 50 delegated and 70 solo runs ($6.45 compared with $11.99 per run in expectation); delegated figures carry a measurement band of about 20%.
5. Agent-architecture scaling: Kim et al., "Towards a Science of Scaling Agent Systems," arXiv.08296, 2025. Independent external study, cited only for the direction of the finding on when delegation does not pay, not for any figure.
6. DeepWideSearch: "DeepWideSearch: Benchmarking Depth and Width in Agentic Information Seeking," arXiv.20168, 2025. The 220 questions span 15 domains, each combining many-row collection with multi-hop retrieval; measured on the benchmark's standing row set, 3 runs per configuration.
7. DeepResearch Bench II: Li et al., "DeepResearch Bench II: Diagnosing Deep Research Agents via Rubrics from Expert Report," arXiv.08536, 2026. Its 132 research tasks across 22 domains are graded against expert-derived binary rubrics; measured on a 50-task subset stratified across all themes, one attempt per task, 3 runs, scored on tasks no configuration refused. Claude Opus 4.6 judges under the benchmark's rubric protocol; the original uses a different judge, and an Anthropic judge may favor the house style. The runs predate Claude Opus 5, hence its absence from the Compare models chart. The Sonnet 5 cost differs slightly between the caching chart (inference cost, with and without caching) and that chart (all-in cost, caching on); both come from the same runs.
8. Corpus defect sweep: Anthropic-internal, for work larger than one context window: a 21.6-million-token corpus from 14 public Python package sources with 130 planted defects and deterministic grading; protocol fixed before the runs and internally reviewed; three runs per configuration. The charted team configuration ran its workers on Claude Managed Agents: each of six waves was one session in which the Claude Fable 5 lead ran 40 Claude Sonnet 5 worker threads, the platform's limit at the time (the current default is lower), with a small external driver sequencing the waves and carrying the findings between them. Absolute F1 is specific to this corpus build, not comparable across benchmarks; configuration comparisons are like for like.
9. GPQA Diamond: Rein et al., "GPQA: A Graduate-Level Google-Proof Q&A Benchmark," 2023. The 198-question Diamond subset, measured August 2026, two runs per configuration, model-graded against reference answers, advisor tokens metered per request. A platform safety check refused two biology questions on the Sonnet and Opus executors; excluding them changes no comparison by more than one point.
10. DeepSWE: Datacurve, "DeepSWE: Measuring Frontier Coding Agents on Original, Long-Horizon Engineering Tasks," arXiv.07946, 2026. Measured August 2026: 113 original tasks across five languages with program-based verifiers. Pairings are two runs each with advisor tokens metered per request, and used a client-side advisor loop rather than the advisor tool, with identical accounting. Single-model effort sweeps are single runs priced from token counts, a cache-aware approximation. Costs per task are run totals divided by 113.
11. Internal agentic-coding benchmark: Anthropic-internal: 370 repository tasks graded by the repositories' own tests. The API figures (Opus 5 alone, Fable 5 alone, and the pairing) were measured August 2026 at the default effort with a 128,000-token output cap, one run per configuration: five attempts per task at the default settings and for the pairing, one at low and medium; the pairing averaged about two advisor consultations per attempt; costs are per attempt. The Claude Code figures are July 2026 runs of the same tasks, one run per configuration, costs approximate.
12. Internal repository-task benchmark (cap measurement): A separate Anthropic-internal set of about 130 repository tasks, run August 2026 with a plain API agent loop, one attempt per task. The 16,384-token figures average two runs per model; the 64,000-token figures are single runs (124 tasks scored for Opus 5; 108 for Claude Fable 5, the environment having skipped the rest before the model ran). About half the Fable attempts the 16,384 cap had ended solved at 64,000; a further Fable run at 128,000 scored 56.1%, within noise of the 64,000 run. The SWE-bench Pro cap figures are one Claude Fable 5 run per cap at the default effort on a 100-problem subset stratified from reference 3's 482-problem set, not comparable to its scores. The chart's per-turn distributions come from the Opus run at 64,000 and the Fable run at 128,000, so neither is cut off by its own cap.
13. Chartography: Surge AI, "Chartography," 2026. The complete released 100-question set, measured August 2026 with Anthropic's implementation on Claude Managed Agents (standard cloud sandbox; advisor configurations use the Managed Agents advisor). Claude Sonnet 4.6 grades instead of the reference judge and the benchmark runs with tools, so scores compare across configurations here but not to the published leaderboard. Two runs per configuration, pooled; run-to-run spreads were 4 to 10 points. Costs exclude sandbox time, which added under 1%. The consult-rate comparison comes from rerunning the same configurations on the Messages API with a container tool set.
14. Support-desk prompt-audit evaluation: An Anthropic-constructed set of 44 support tickets with deterministic grading, run August 2026 under six system prompts, each adding to the same clean prompt one pattern common in prompts written for Claude Opus 4.8 and Claude Sonnet 4.6. Each chart point is one of three cases (older model, newer model on the same prompt, newer model after the audit) averaged over the six prompts and 44 tickets. The Opus 5 accuracy gain has a 95% confidence interval of 3 to 8 points; the Sonnet accuracy differences are within noise.

Prompt caching

The largest free win on this page: setup, lifetimes, and diagnostics.

Effort

Trade intelligence for latency and cost within a single model.

Choosing the right model

Evaluate capability, speed, and cost across the Claude model family.

Task budgets

Give agent loops a token countdown they self-regulate against.

Session budgets

Put a hard dollar cap on a Managed Agents session.

See current per-token pricing for every Claude model.

Watch a walkthrough of Claude Fable 5 and the advisor and orchestrator patterns.

Was this page helpful?