Cost optimisation · 2026 · MeasuredSheet GP-COST-01 · Unit of account

Cost per task,
not per token

Anthropic's official documentation attaches a measured result to every cost lever. Once the unit of account becomes cost per completed task, several widely used ways to save money reverse: where outcomes are checkable, running everything at a low effort setting and re-running only the failures reaches the same pass rate for about half the cost; multi-model architectures saved money in only two measured situations; and lowering max_tokens saves nothing.

SourceClaude Platform Docs, "Optimizing for cost and intelligence" MeasuredAnthropic-internal runs · July to August 2026 CheckedEvery figure verified · all 19 source charts read directly Date2026-08 · v1
Priced per token Priced per completed task costlier cheaper Fable 5 · ~5x per token low, $0.76 Sonnet 5 · baseline default, $0.84 Opus 5 · ~2.5x runs predate it DeepResearch Bench II, 50-task subset: the ranking reverses

GenAI Playbook · a faithful reading of the official documentation · all measured figures are Anthropic-internal runs, and the page itself states they are directional, not guarantees

Group one · quality preserved

The documentation puts the free wins first

Anthropic's official platform documentation divides its cost controls into two groups: free wins that preserve quality, and tradeoffs that exchange cost for intelligence. Most of what it measures is an agent loop, where a model works through many turns with tools, though a few of the benchmarks are question answering. The page ranks prompt caching first by a wide margin.

A token is the unit used for metering and billing, roughly a word or half a word, with separate input and output prices. Every turn resends the system prompt, tool definitions, and all prior messages. A 40-turn task sends its first turn 40 times, so total task cost grows with roughly the square of turn count.

Caching does not stop this repetition. It changes its price. The cached material must form a prefix, the request content beginning at the first token, and matching is byte-for-byte, so a change invalidates that point and everything after it. A cache read, meaning a hit on stored content, is billed at 0.1x the input price. A five-minute cache write, which stores new content, is billed at 1.25x; a one-hour cache write costs 2x. The default cache lifetime is 5 minutes, and an agent loop's turns are seconds apart, so the discount applies to most tokens on every turn.

0.1x
A cache read on a matched prefix bills at a tenth of the input price
3.7x
On WideSearch, caching made each problem 3.7x cheaper
12%
Caching plus input trimming left the triage bill at 12% of baseline
50%
The Batch API discounts every token by 50%, cached ones included
$0$5$10 $15$20 Claude Fable 5 · WideSearch Claude Fable 5 · DeepResearch Bench II Claude Sonnet 5 · DeepResearch Bench II $5.5 $20 3.7x cheaper $2.4 $6.0 2.5x cheaper $0.80 $2.8 3.4x cheaper No caching With caching
Figure 1 · Cost per problem, re-priced from identical per-request billing records · the charted runs achieved hit rates of 81% to 90% · WideSearch measures broad web research through the completeness and accuracy of a many-row table; DeepResearch Bench II grades research reports against expert-derived rubrics · redrawn from the official chart

The issue-triage experiment makes the mechanism concrete. The model was claude-sonnet-5, running on a frozen snapshot of a large public repository and working through 20 real bug reports containing screenshots. Caching alone cut 83%, and caching plus input trimming reached 88%.

A longer variant processed 2.6x the tokens: its baseline cost $8.56, caching reduced that to $1.34, and one compaction event brought it to $0.83, a further 38%. Compaction summarizes earlier conversation once it grows too long so the loop can continue with a shorter history, and it needs a session long enough to trigger. The 20-issue run never reached the 50,000-token floor once its inputs were trimmed.

20-issue run (baseline $3.32) Longer run, 2.6x the tokens (baseline $8.56) 0$1$2$3 $3.32$0.58$0.39 17% of baseline12% of baseline baseline + prompt caching + input trimming image resizing · tool search 0$0.4$0.8 $1.2$1.6 $1.34$0.83 38% less prompt caching + compaction fired once
Figure 2 · The same triage job on a frozen snapshot of a large public repository · one run per bar, cumulative within a panel · the panels use different scales · redrawn from the official chart

Three mid-task changes break the cache: changing effort, changing a task budget, and running context editing. The documentation puts all three at natural boundaries, followed by confirmation that cache reads have not dropped, and it sets a task budget once on the first request rather than changing it later. The request-structure design behind stable caching is covered elsewhere on this site and is not repeated here.

The Batch API is the next free lever for work that does not require an immediate response. It discounts every token by 50%, including cached tokens, with results delivered any time within 24 hours. It is unavailable for sessions on Claude Managed Agents, Anthropic's hosted agent runtime where sessions, sandboxes and billing sit on the platform side, because those sessions are interactive by design.

Nominally free, measured otherwise

Context editing clears stale tool results to make room in the context window, but each clearing pass rewrites the cached conversation and works against prompt caching. In the run measured, it cost more than it saved. Its purpose is making room, not saving money, and the documentation favors a few large clearing passes over many small ones.

Same group · easily missed

A prompt written for a previous model is a standing tax

Each model generation responds to prompts differently, so a prompt accumulates instructions added for a model no longer in use, typically "verify twice," "be maximally thorough," a mandatory step-by-step procedure, or a hand-rolled reasoning scratchpad. A newer model follows those to the letter, adding tool rounds and output without improving accuracy.

Anthropic measured this on a support-desk evaluation: 44 tickets, six prompts written for the older models, deterministic grading, with accuracy movements under about 5 points treated as noise.

ConfigurationAccuracyCost per ticketWhat the step bought
Claude Opus 4.8 + old prompt89.5%2.52 centsstarting point
Claude Opus 5 + the same old prompt91.8%3.43 cents36% more no change in accuracy
Claude Opus 5 + audited prompt97.0%2.93 cents14% cheaper accuracy rose to 97%
Claude Sonnet 4.6 + old prompt81.1%1.72 centsstarting point
Claude Sonnet 5 + the same old prompt83.9%1.61 centsaccuracy difference within noise
Claude Sonnet 5 + audited prompt85.6%1.39 cents14% less same accuracy
1.01.52.0 2.53.03.5 808590 95100 cost per ticket (cents) tickets resolved correctly (%) Claude Opus 4.8 · old prompt Claude Opus 5 · same old prompt Claude Opus 5 · audited prompt Claude Sonnet 4.6 · old prompt Claude Sonnet 5 · same old prompt Claude Sonnet 5 · audited old model · old prompt new model · same old prompt new model · audited prompt circles = Opus migration · diamonds = Sonnet migration
Figure 3 · Support-desk evaluation, 44 tickets, deterministic grading, each point the mean over prompts and tickets · redrawn from the official chart

Swapping in the new model while keeping the old prompt cost 36% more per ticket than the old model on the old prompt, for no change in accuracy. After the audit, Claude Opus 5 was 14% cheaper than its unaudited version and accuracy rose from 92% to 97%, a gain outside the noise with a 95% confidence interval of 3 to 8 points. On the Claude Sonnet 4.6 to Claude Sonnet 5 migration the audit removed 14% of the cost at the same accuracy.

The two kinds of stale text have different costs, and the per-pattern measurements support that split: instructions the new model follows too literally cost money, while text that no longer fits the model costs accuracy.

Cost change (cents per ticket, negative is a saving) Accuracy change (points, positive is a gain) inside the band = under about 5 points, within noise -1.5-1.0-0.50+0.5 -50+5+10+15 "Verify twice before submitting" "Be maximally thorough" Mandatory six-step procedure Hand-rolled scratchpad reasoning Contradictory rules Retired thinking setting −1.43 −0.44 −1.19 −0.12 −0.53 −0.72 −0.31 −0.23 +0.20 +0.01 +0.31 +0.10 −2.3 −2.3 +6.8 +1.1 +0.0 +2.3 +9.1 −5.7 +6.8 +13.6 +11.4 +1.1 Opus 5 Sonnet 5
Figure 4 · Each prompt carries one legacy pattern · bars are each post-audit result minus its pre-audit result, averaged over 44 tickets · redrawn from the official chart

Removing "verify twice" cut Claude Opus 5's cost per ticket by a third, and removing "be maximally thorough" almost as much. The retired thinking setting, the contradictory rules, and the hand-rolled scratchpad each restored 7 to 11 accuracy points on Claude Opus 5 when removed. The documentation notes that the same patterns appear in tool descriptions and skills, and are worth removing there too.

With the free levers exhausted, every remaining control trades quality for cost.

A new unit of account

The unit of account changes the ranking

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. What is paid for, though, is completed tasks rather than tokens. A more capable model finishes a task with less work: fewer turns, less searching, less re-reading of its own context, less backtracking. That reduction routinely overwhelms the per-token premium.

Anthropic measured this directly on a research-report benchmark. The DeepResearch Bench II 50-task subset was run three times and graded by Claude Opus 4.6.

ConfigurationRubric scoreCost per completed task
Claude Fable 5 · low effort60.2$0.76
Claude Sonnet 5 · default56.0$0.84

The frontier model at low effort was more accurate and about 10% cheaper per task than the mid-tier model at its default, despite about 5x the per-token price. These runs predate Claude Opus 5, so this comparison contains no Claude Opus 5.

The reversal is not universal. On this page's SWE-bench Pro subset, where models edit real repositories and must pass their tests, Claude Opus 5 alone scored 91.7% and Claude Fable 5 alone 91.3%, a difference inside run-to-run noise, and Claude Opus 5 cost about 60% as much. The documentation therefore recommends Claude Opus 5 as the starting point for most agent workloads rather than the frontier model: per token it is half of Claude Fable 5 and 2.5 times Claude Sonnet 5, and on that coding subset it matched Claude Fable 5's accuracy.

The ranking flips by workload and no price list says which way, so every candidate has to be priced in cost per completed task on real traffic, including Claude Opus 5 and the frontier model at reduced effort.

One more thing the documentation changes with the unit

The documentation puts pricing on the hardest tenth of the workload, not the typical task. 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.

Even with nothing failing, spending concentrates in the tail. Across 20 WideSearch problems run three times with Claude Fable 5, total cost was $421.

$0 $20 $40 $60 $80 $100 $120 1 5 10 15 20 20 WideSearch problems, ranked by cost · Claude Fable 5 cost ($, summed over three runs) the two most expensive problems (10%) carry 43% of all spend the single most expensive one carries 30% the cheapest half together: 10%
Figure 5 · Per-episode billing records, $421 across three runs · redrawn from the official chart

The first question is therefore not which model, but where the effort setting on one model should sit.

Inside one model

The effort dial has two shapes

effort is the most direct cost control inside a single model. It governs how much thinking, tool calling and self-verification the model performs at low, medium or high, and the default is served as high rather than as a separate setting. Cost rises with all of that activity, while accuracy rises only with the part a task needs. Below the model's capability ceiling, the highest settings pay for depth the task never uses.

The sweep produced two curve shapes, with opposite conclusions. The first is nearly flat, across research and knowledge work; the second, on long-horizon coding, is steep.

WideSearch (Claude Fable 5) 74 76 78 80 82 84 3 4 5 6 cost per completed task ($) low medium default 3 runs per point DeepWideSearch (Claude Fable 5) 60 62 64 66 68 70 3 4 5 6 cost per completed task ($) low medium default orchestrator + Sonnet 5 worker 3 runs per point BrowseComp (Claude Fable 5) 74 76 78 80 82 84 4.0 4.5 5.0 5.5 6.0 6.5 cost per completed task ($) low medium default chart says 3 runs; reference 4 says one to three GDPval (Claude Fable 5) 78 80 82 84 86 88 1.0 1.5 2.0 2.5 3.0 3.5 cost per completed task ($) low medium default 1 run per point, cost includes grading SWE-bench Pro (Claude Opus 5) 80 85 90 95 0.4 0.8 1.2 1.6 cost per completed task ($) low medium default default = 2 runs, others 1 accuracy (%) Two shapes Research and knowledge work (top row and GDPval): the line is nearly flat, so the lower settings buy most of the accuracy for a third to a half of the cost. Long-horizon coding (SWE-bench Pro): the line is steep, so the top of the dial is paying for accuracy the task actually uses. Every point priced from metered tokens at current prices; 'default' is served as high effort. Panels use their own scales.
Figure 6 · Each panel is one model on one benchmark, the same tasks at three effort settings · no y-axis starts at zero · redrawn from the official chart
Benchmarklowmediumdefault
WideSearch
broad web research into a many-row table
78.5% / about $2.979.8% / about $4.280.0% / about $5.9
DeepWideSearch
many-row collection plus multi-hop retrieval
64.9% / about $3.465.9% / about $4.666.4% / about $6.1
BrowseComp
whether browsing agents find hard-to-find facts
78.5% / about $4.281.5% / about $5.381.3% / about $6.15
GDPval
economically valuable knowledge-work deliverables
83.5% / about $1.4584.0% / about $2.2583.1% / about $3.2

All four swept with Claude Fable 5. WideSearch and DeepWideSearch are 3 runs per point; BrowseComp's chart says 3 runs while reference 4 says one to three runs per setting; GDPval is 1 run per point and its cost includes grading.

All four point the same way: low gives up 1 to 3 points for a third to a half of the cost per task, medium matches the default's accuracy at 70% to 85% of its cost, and the default buys nothing measurable over medium on any of the four. On BrowseComp and GDPval the default is not more accurate than medium, and on the chart its score sits slightly lower.

Lower settings are also faster, which matters when latency is the constraint. DeepWideSearch takes 4.5 minutes per problem at low against 7.9 minutes at the default, and on the corpus benchmark an episode takes 7.9, 9.1 and 11.4 hours at low, medium and the default. Anthropic also reports that low on DeepWideSearch matched an orchestrator using a Claude Sonnet 5 worker at 20% lower cost, so lowering effort beat an architecture change.

The second shape appears on long-horizon coding. On the SWE-bench Pro 482-problem subset, Claude Opus 5 at medium gives up about 2 points for half the cost and at low about 8 points for a quarter of it. That is a real tradeoff, not a free win. One parameter behaves categorically differently across the two kinds of work.

Some workloads reach the model's ceiling, and there every step buys score.

586062 6466 0.51.01.5 2.02.5 cost per task ($, list) rubric score (%) low · 60.2 / $0.77 medium · 62.7 / $1.48 high (default) · 65.1 / $2.38 about 2.4 points per step this curve offers no free cost reduction
Figure 7 · Claude Fable 5, the DeepResearch Bench II 50-task subset, 3 runs per setting · y-axis starts at 58 · redrawn from the official chart

The task description alone does not identify which curve applies, so the method is to sweep two or three settings on a sample of real traffic and read the answer off the curve. Each setting needs its own session, because changing effort mid-session invalidates the cache and distorts the comparison.

Two consequences follow in the documentation. 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. And this single-model curve is the baseline any multi-model strategy has to beat.

How far the work may go

The three output-side controls

effort governs how much work goes into a task; the output side governs how far that work is allowed to continue. Re-running failures, task budgets and max_tokens look like one family, but only the first two save money, and the dividing line is whether the model can see the constraint.

When outcomes are automatically checkable, a fixed setting is itself waste

The cheapest policy on the effort curve is not a fixed setting but running everything low and re-running only the failures higher. Claude Opus 5's points on the same SWE-bench Pro tasks are below.

828486 88909294 0.20.40.6 0.81.01.2 1.41.6 cost per task ($, incl. the failed cheap attempt) pass@1 (%) low alone · 84.0% / $0.37 medium alone · 89.4% / $0.72 default alone · 91.7% / $1.39 low throughout, failures re-run at default 93.1% / $0.70 medium throughout, failures re-run at default 94.0% / $0.95 fixed setting alone (small dots = the default's two runs) run low or medium throughout, re-run only failures
Figure 8 · Claude Opus 5, the SWE-bench Pro 482-problem subset, computed task by task from the effort-ladder runs · y-axis starts at 82 · redrawn from the official chart

16% of tasks failed at low. Re-running those at the default reached the default's pass rate at about half the price per task. The documentation adds one qualification: most of the small lift comes from having a second attempt rather than from the higher setting, because re-running the default's own failures at the default also reached 94.0% but cost $1.58. Use it for the saving, not the lift.

The second condition is latency: every first-pass failure consumes two runs of wall-clock time, so the saving is paid for in latency on the failures.

A task budget saves money because the model can see it

Most agentic runs are inexpensive, while a minority spend many times the median on additional searches, repeated verification and excessive testing, and the budget targets that tail. With a live token countdown for the whole task, the model trims low-value searches, skips redundant verification, and wraps up instead of spiraling.

86889092 1.01.41.82.2 cost per task ($, list) pass@1 (%) 20,000 tokens 35,000 tokens no budget tightest budget: −4.4 pts for −47% cost output tokens fall 41% to 64% down the ladder
Figure 9 · Claude Fable 5, SWE-bench Pro, 482 problems fully scored per budget level · the source states the run count two different ways; see the closing methodology notes · redrawn from the official chart

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

The documentation suggests starting near the loop's 90th-percentile token usage and then tightening. Budgets below the current 20,000-token floor are rejected, and very tight budgets can produce refusal-like behaviour. The budget is set once on the first request, because a mid-task change invalidates the cache. And it is advisory, steering the model rather than stopping it, so adherence has to be verified on the actual workload.

Lowering max_tokens saves nothing, because the model cannot see it

It 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.

Claude Opus 5: solved 53% at the 16k cap, 58% at 64k Claude Fable 5: solved 33% at the 16k cap, 55% at 64k $0$10$20 $30$40$50 $18 $19 $34 $33 cost per attemptcost per solved task $17 $29 $53 $53 cost per attemptcost per solved task max_tokens 16,384 max_tokens 64,000 none of the attempts the 16k cap ended were solved
Figure 10 · An internal set of about 130 repository tasks, August 2026, a plain API agent loop, one attempt per task · the 16k-cap bars average two capped runs, the 64k-cap bars are single runs · redrawn from the official chart

The 16,384 cap ended 15% of Claude Opus 5's attempts and a third of Claude Fable 5's, and none of the truncated attempts were solved. The lower cap did reduce spending per attempt, but it solved proportionally fewer tasks, so cost per solved task was the same at both caps. At 64,000 nothing was cut off and Claude Fable 5 solved 54.6% instead of 33.3%; on a separate cut of the SWE-bench Pro subset it solved 92% instead of 90%.

The turn-length distribution explains how a seemingly generous cap makes that much difference.

1001,000 10,000100,000 output tokens in one turn (log scale) 16k cap 32k 64k 128k max Claude Opus 5 Claude Fable 5 median245 90th1,894 99th8,813 longest turn33,363 0.22% of turns exceeded 16,384 tokens; that cap ended 15% of attempts median293 90th2,134 99th13,005 longest turn59,499 0.55% of turns exceeded 16,384 tokens; that cap ended a third of attempts
Figure 11 · The same internal repository tasks · the Claude Opus 5 run used a 64,000-token cap and the Claude Fable 5 run 128,000, and no turn in either run reached its own cap · redrawn from the official chart

A very small tail of long turns decides a large share of outcomes. Retrying a capped attempt only adds cost: at the same cap those attempts never succeeded, and at a higher one the wasted attempt is still on the bill. Anthropic's recommendation is to set max_tokens to 64,000 for agentic work and 128,000 at xhigh or max effort, stream responses that large, treat stop_reason: max_tokens as a failure, and leave saving to the two controls the model can see.

ControlVisible to the modelWhat it doesOn hitting the limit
Task budget
advisory, beta
visiblethe model converges on its own, so it saves moneysteers rather than stops; adherence has to be verified
max_tokens
single-response cap
invisiblea safety cap against an unbounded single response, and saves nothingreturns stop_reason: max_tokens, treat as a failure
Session budget
Claude Managed Agents
platform-enforceda hard dollar stop at list rates for tokens, searches and session timepauses with stop_reason: budget_reached, resumes when raised
What the documentation recommends

A session budget works on any model with a list price including Claude Sonnet 5, and combines with the advisory task budget. The documentation recommends setting all three with a workspace spend limit as the final backstop. All of these controls still operate inside one model. A second model enters only when tuning them is not enough.

A second model

When a second model is worth adding

Multi-model architecture is a common answer to cost, and Anthropic's own measurements support a more conservative reading. In the documentation's account, the approach fits workloads with uneven difficulty: routine steps stay with a lower-cost model while a frontier model handles the decisions that need more capability, which keeps most tokens billing at smaller-model rates. If difficulty is uniform, or the work forms one dependent chain, a well-tuned single model is usually the stronger baseline.

Two architectures are measured, distinguished by which model holds the main loop. In the advisor strategy a lower-cost executor runs the loop, asks a higher-intelligence advisor for guidance at difficult decisions, and then continues; the advisor tool provides this at the request level, with the executor emitting a tool call, Anthropic running the advisor inference server-side, and the advice returned without any orchestration code. In the orchestrator strategy the frontier model retains the loop, divides the task, sends partitions to lower-cost workers, and merges their results.

Advisor strategy · the lower-cost model holds the loop Main loop Executor Claude Sonnet 5 runs every turn Tool call Sends advice Advisor Claude Fable 5 On demand The stronger model is billed only for the consultations; the capability gap and the consult rate decide the outcome. Orchestrator strategy · the frontier model holds the loop Main loop Orchestrator Claude Fable 5 Plan and merge Fan out worker 1 Claude Sonnet 5 worker 2 Claude Sonnet 5 worker 3 Claude Sonnet 5 Most tokens bill at worker rates, and the price is a plan, a handoff and a merge.
Figure 12 · The two multi-model shapes in the documentation, distinguished by which model holds the main loop · redrawn from the two official diagrams

The advisor strategy: the gain tracks the consult rate

Two variables decide how much an advisor helps. The first is the capability gap, since an advisor can supply only capability the executor lacks: on GPQA Diamond 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 is more fragile: whether the executor actually asks. The consult rate is the share of tasks on which it calls the advisor, and across six pairings the realised gain follows it closely.

0 10 20 30 -5 accuracy points Claude Haiku 4.5 executor + Claude Opus 5 advisor · GPQA Diamond +26 pts of 29 available asked on 90% of tasks Claude Opus 5 at low effort + Claude Fable 5 advisor · chart reading +18 pts of 25 available asked on 86% of tasks Claude Sonnet 5 executor + Claude Opus 5 advisor · SWE-bench Pro +9.5 pts of 16 available asked on 73% of tasks Claude Sonnet 5 executor + Claude Opus 5 advisor · GPQA Diamond +4.5 pts of 5 available asked on 37% of tasks Claude Opus 5 executor + Claude Fable 5 advisor · coding tasks +1.3 pt; almost nothing available asked on 100% of tasks Claude Sonnet 5 at low effort + Claude Opus 5 advisor · SWE-bench Pro −5.4 pts of 16 available asked on 7% of tasks; scored below the executor alone gap available (advisor's model alone minus executor alone) gain the pairing realised
Figure 13 · Each row is one pairing measured for this page · GPQA and chart reading are two runs per configuration; the run counts for SWE-bench Pro and the coding tasks are stated two different ways in the source; see the closing methodology notes · redrawn from the official chart

When the executor did ask, the advisor closed 60% to 90% of the gap to the stronger model while that model was billed only for the consultations, which is what makes the cost cases possible. The last row is the counter-case: the same low-effort executor asked on only 7% of tasks and the pairing scored 5.4 points below the executor alone, because an executor at low effort can stop noticing it is stuck.

Prompting moves the consult rate. With only the tool's built-in description executors under-call, particularly on coding work, and the documented system prompt asks for one call before substantive work and one before finishing, about two to three calls per task. The rate has to be prompted for, measured, and watched.

On cost, the documentation is notably restrained. A few short consultations replace running the advisor's model for the whole task only when that model is priced well above the executor, which makes a frontier advisor over a mid-tier executor the most cost-effective shape. Two measured sets follow.

Internal coding benchmark (370 repository tasks)SolvedPer attempt
Claude Opus 5 executor + Claude Fable 5 advisor85.7%$8.40
Claude Opus 5 alone, default84.4%$8.50
Claude Fable 5 alone, medium83.4%$8.20
Claude Opus 5 alone, low74.3%$1.9
Claude Opus 5 alone, medium82.1%$4.5
Claude Fable 5 alone, low78.7%$5.4
Claude Fable 5 alone, default85.1%$11.9

A plain API agent, August 2026. Five attempts per task at the default settings and for the pairing, one at the reduced settings. The pairing averaged about two advisor consultations per attempt. Differences of a point or two are within run-to-run noise.

Chartography (chart reading)lowmediumdefault
Claude Opus 5 alone49 / $0.3875 / $0.9479 / $1.95
Claude Fable 5 alone55 / $0.8173 / $1.4474 / $1.74
Claude Opus 5 at low effort + Claude Fable 5 advisor67.5 / $0.60 (its two runs scored 65 for $0.47 and 70 for $0.73)

The complete released 100-question set from Surge AI, August 2026, on Claude Managed Agents, two runs per configuration pooled, with run-to-run spreads of 4 to 10 points.

Chartography · chart reading (score / cost per task) Internal coding benchmark (share solved / cost per attempt) $0.00 $0.50 $1.00 $1.50 $2.00 $0 $4 $8 $12 45 55 65 75 85 72 76 80 84 88 cost per task ($, list) cost per attempt ($) low medium default low medium default low medium default low medium default pairing 67.5 / $0.60 pairing 85.7% / $8.40 Claude Opus 5 alone (by effort) Claude Fable 5 alone (by effort) executor + advisor pairing (small dots: the pairing's runs)
Figure 14 · Left, the Chartography chart-reading benchmark (Claude Opus 5 at low effort with a Claude Fable 5 advisor); right, the internal coding benchmark (Claude Opus 5 executor with a Claude Fable 5 advisor) · the panels use different scales · redrawn from the official charts

The pairing was the most accurate configuration measured, but it sits only a point or two above the best of each model's own settings, which a single run does not separate from noise. The documentation asks that the result be read as a shape to test against one's own workload rather than as a saving, and places the cost case with pairings that have a wider capability gap.

That wider gap appears on chart reading, where the advisor beat raising the executor's own effort. On Chartography, low-effort Claude Opus 5 with a Claude Fable 5 advisor scored 67.5 for $0.60 per task, above the line through either model's own effort settings, where Claude Opus 5 alone moves from 49 at $0.38 to 75 at $0.94. The executor's own medium and default settings still hold the top scores, at 1.6 and 3.3 times the pairing's price. The low-effort executor in this pairing consulted on 86% of tasks, the condition the SWE-bench Pro pairing failed to meet.

In the documentation's assessment

The advisor strategy suits work whose turns are mostly mechanical but whose plan matters, including coding agents, computer use, and multistep research pipelines. It fits poorly when every turn genuinely needs frontier capability, when there is nothing to plan, or when the executor already approaches the advisor's capability.

The orchestrator strategy: the saving lives in the routine tail

The orchestrator saved money in two measured situations. The first is insurance against the cost tail on routine work. A frontier model running alone occasionally spirals on a routine problem it would normally solve, and because those runs cannot be identified in advance, a few of them dominate the bill. Handing routine work to a lower-cost worker caps that tail, because any spiraling now happens at worker rates.

$0$20$40 $60$80 cost per run ($) · pair right in 50 of 50 runs, solo in 68 of 70 cheaper is left Claude Fable 5 coordinator + one Claude Sonnet 5 worker Claude Fable 5 alone $6.45 expected $12 90th pct $42 worst run $11.99 expected $33 90th pct $84 worst run solo's most expensive run ($84) was also a wrong answer
Figure 15 · Ten BrowseComp problems the solo model reliably solved, drawn from a 26-problem slice; 50 delegated runs (5 per problem) and 70 solo runs; delegated costs carry a measurement band of about 20% · redrawn from the official chart

Anthropic describes delegation as costing a little under half of the solo model on average and about a third at the 90th percentile; by the expected values in the table, $6.45 against $11.99 is about 54% of solo, which is a little over half. The counter-intuitive part, as the documentation puts it, is that delegation paid on the routine, normally solvable share of the work, the opposite of the intuition that workers are for hard problems. The boundary belongs in the same breath: on the full, harder BrowseComp set the economics reversed.

The second is work larger than one context window, the amount a model can read in one pass. A solo model must process such material serially, paying to re-read its own state on every pass, while workers each read their own partition in parallel at worker rates. Anthropic built a benchmark for this case: a 21.6-million-token corpus drawn from 14 public Python package sources with 130 planted defects and deterministic grading.

$200$300$400 $500$600$700 $800 0.700.750.80 0.850.90 all-in cost per episode ($) planted defects found (F1 score) low · 0.775 / about $740 medium · 0.839 / about $720 high (default) · 0.877 / about $764 Claude Sonnet 5 alone · 0.762 / about $228 Claude Fable 5 lead + Claude Sonnet 5 worker 0.808 / about $332 55% cheaper than every Fable 5 solo setting; 3 to 7 points below medium and default Solo cost barely moves with effort ($720 to $764, inside seed noise): the bill is reading the 21.6M-token corpus, so on this workload the effort dial only changes accuracy. Claude Fable 5 solo (by effort) Claude Sonnet 5 solo Fable 5 lead + waves of Sonnet 5 workers
Figure 16 · A 21.6-million-token public-Python-source corpus with 130 planted defects and deterministic grading · 3 runs per configuration, with the protocol fixed before the runs and internally reviewed · y-axis starts at 0.70 · redrawn from the official chart

Lowering effort cannot help here, because the bill is reading the 21.6-million-token corpus itself: Claude Fable 5 alone 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 Claude Fable 5 solo settings and scored 3 to 7 points below Claude Fable 5 at medium or the default, while beating a Claude Sonnet 5 solo baseline outright.

The token accounting shows why. The coordinator configuration read 32 million input tokens against the solo model's 14.5 million and still cost less, because partitioned reading at worker rates is cheaper than repeated re-reading at frontier rates. Claude Fable 5 at the default still holds peak accuracy, at 2.3 times the coordinator configuration's cost, so delegation here buys most of the accuracy rather than all of it.

F1 balances completeness against precision in finding the planted defects, and its absolute value is specific to this corpus build.

When delegation does not pay

An orchestrator buys something only when there is genuine bulk to hand off, ideally more than one context window of it. For one dependent chain, or work that fits in a single context, it pays for a plan, a handoff and a merge that a single model gets for free. On the full, harder BrowseComp set the frontier model alone reached the coordinator configuration's accuracy at 22% to 30% lower cost. Across every case of this kind that was measured, the coordinator's model running alone at lower effort came out ahead. The documentation also references an external study on scaling agent systems, cited only for the direction of the finding and not for any figure.

The order of evaluation follows from that evidence. Sweep the current model's effort settings first, the cheapest experiment on the page, where most workloads end. If a gap remains, price the stronger model alone at low effort, since that is the number a pairing has to beat, and the pairings on this page that beat it were the ones whose executor actually consulted. On adoption cost, the documentation notes that adding an advisor is a tool definition rather than a rearchitecture.

On real traffic

Measuring it on real traffic

These readings come from July and August 2026 at the list prices then in effect, and they will drift as models and prices change. Escalation rates, how cleanly tasks split, and transcript length move them too. The documentation's method stays the same, in four steps.

Sample real traffic and define an outcome check

Pull tasks from production logs in proportions that reflect real traffic, write an outcome check for each (tests pass, ticket closed, row count correct), and record cost per task beside the score.

Baseline the model tiers across effort levels

Not only the default. Plot score against spend. A multi-model configuration has to beat the single model's whole curve.

Add architecture only where the curve leaves a gap

If the curve exposes a gap that effort cannot close, add the multi-model strategy that fits and re-run the suite.

Shadow the winner, then keep the suite running

Run the winner in shadow on a traffic slice before cutover, then keep the suite running.

Cost per task is the sum across the task's requests of the four token counts reported in each response's usage, each at its own rate: input at the input price, five-minute cache writes at 1.25x the input price, cache reads at 0.10x, and output at the output price. The worked example uses Claude Opus 5 list prices.

python # Per-million-token prices from the pricing page; # change these two for another model. INPUT_PER_MTOK = 5.0 # Claude Opus 5 OUTPUT_PER_MTOK = 25.0 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
One self-check worth running

The documentation adds one self-check: in an agent loop the cache-read term is usually the largest of the four, and if it is not, caching should be checked.

One thing easy to miss

With the advisor tool or compaction enabled, some tokens are reported only in usage.iterations and not in the top-level totals, so the sum has to run over usage.iterations, with advisor_message entries priced at the advisor model's rates.

The page closes with a table of levers in the order to try them, ordered free wins first, tradeoffs next, multi-model last. The readings will drift; the method will not.

Provenance

Methodology and provenance

All measured results are Anthropic-internal runs of these benchmarks, not third-party reproductions and not invoice amounts; charts labelled notional USD price each request's token counts at list rates. Unless otherwise noted, costs are at August 2026 list prices, with Claude Sonnet 5 at $2 and $10 per million input and output tokens. The page itself states the results are directional and not guarantees.

BenchmarkComparability limit
SWE-bench ProUses a 482-problem subset selected for compatibility with Anthropic's evaluation harness, and its scores are not comparable to the public leaderboard; the max_tokens figures use a further 100-problem subset stratified from it, whose scores are not comparable to the 482-problem set either
ChartographyGraded by Claude Sonnet 4.6 and run with tools, so its scores compare configurations on this page but not to the published leaderboard
GDPvalGraded by a Claude model, so absolute scores may differ from published results
DeepResearch Bench IIGraded by Claude Opus 4.6 where the original benchmark uses a different judge, and an Anthropic judge may favour the house style; those runs predate Claude Opus 5, which is why the model comparison contains no Claude Opus 5
The corpus benchmarkIts absolute F1 is specific to that corpus build

Run counts. On SWE-bench Pro, Claude Opus 5 at the default effort is the mean of two runs while the reduced settings are single runs; the max_tokens 64,000 figures are single runs; the internal coding benchmark is one run per configuration; GDPval is one run per point.

The source contradicts itself in two places on run counts, and both readings are recorded rather than reconciled

Both are official text and there is no way to tell which governs, so neither is adopted as settled. Neither affects any point value.

① The task-budget chart's footnote says 35k = mean of two runs, while reference 3 says the task-budget figures are one run per budget. Reference 3 separately states that every budgeted run completed all 482 problems without harness errors.

② The advisor-mechanism chart's footnote says SWE-bench Pro and the coding tasks are one run per configuration, while reference 3 splits the two SWE-bench Pro pairings: the default-effort pairing was run twice (a run and an exact replication) and the low-effort pairing once.

Noise bands. On the support-desk evaluation, accuracy changes under about 5 points are within noise, the Claude Opus 5 accuracy gain has a 95% confidence interval of 3 to 8 points, and the Claude Sonnet accuracy differences are within noise. On the triage run, differences of about $0.10 are noise. Chartography's run-to-run spreads were 4 to 10 points. On the corpus benchmark, $720 to $764 is inside seed noise. On the internal coding benchmark, differences of a point or two are within run-to-run noise.

Two BrowseComp slices, not interchangeable. The cost-insurance figures use ten problems the solo model reliably solved, drawn from a 26-problem slice, with delegated costs carrying a measurement band of about 20%; the effort figures use a 500-problem cut.

External citation. The external study on scaling agent systems (Kim et al.) is cited for the direction of the delegation finding only, and the source explicitly does not draw any figure from it.

Implementation differences. The DeepSWE pairings used a client-side advisor loop rather than the advisor tool; the Chartography consult-rate comparison came from rerunning the same configurations on the Messages API with a container tool set.