Claude Fable 5.2 pricing, benchmarks and cost per task

Updated November 9, 2026 · first published November 9, 2026

Claude Fable 5.2 launched on November 9, 2026 at $8 per million input tokens and $40 per million output tokens, 20% below Fable 5.1. Repricing our historical Claude Code usage puts it at 1.71× to 1.85× Opus 5.5 on identical tokens. Whether Fable earns that premium depends on how many more tasks it completes. API pricing.

What changed in Claude Fable 5.2?

Fable 5.2 is Anthropic’s broadly available model for demanding coding, reasoning and long-running agent work. The Claude API ID is claude-fable-5-2. It takes text and images, emits text, and offers a 1 million-token context window and up to 128,000 output tokens. Adaptive thinking is always on. The API defaults to high effort, while Claude apps and Claude Code default to medium for this model. Pin effort in a comparison so a difference in defaults doesn’t become a difference attributed to the model. Model specifications · Launch announcement.

Mythos 5.2 shares the underlying model and published token rates. Fable has stronger safeguards for high-risk cyber and biology tasks. Mythos access runs through verification programs. The coding and vision comparisons below use the Fable results, and the public API prices apply to first-party billing. Amazon Bedrock, Google Cloud and Microsoft Foundry are separate purchasing routes with platform-specific availability and billing. Release and availability · Cloud pricing.

How much does Claude Fable 5.2 cost?

All five standard Fable token rates fell 20% from 5.1. For a fixed mix, the token bill falls exactly 20%. At that discount, the new model can consume 25% more of the same mix before the token bill returns to the old level. Against Opus 5.5, uncached input, output and cache writes cost twice as much, while cache reads cost the same. Against Sonnet 5.5, cache reads cost twice as much and the other categories cost four times as much. Prices are global Claude API list rates checked November 9. Anthropic pricing.

Batch costs $4 input and $20 output per million for asynchronous requests. Fast costs $16 input and $80 output, with caching multipliers applied to the higher rates. Fast and Batch are separate modes. Budget server tools, managed runtime, any regional premiums and negotiated pricing separately. Reasoning tokens are billed as output, so the visible answer alone understates output spend. Billing by feature.

For scale, a single standard request with 900,000 uncached input tokens and 20,000 output tokens would cost $8.00. A more typical agent budget spans many requests, each with its own cache hits, writes, new input and output. A large context limit tells you how much fits in a request. Use the token buckets across the entire job to estimate the bill.

What would Fable 5.2 cost on real Claude Code token mixes?

We repriced 22,411 Opus 5.5 turns and 486 Fable 5.1 turns from two published cohorts covering September 11 to October 9, 2026. We multiplied each cohort’s five token totals by each model’s rates and summed them. For Fable 5.2, with token counts in millions, the formula is 8 × input + 40 × output + 0.20 × reads + 10 × 5m writes + 16 × 1h writes. Uncached input and cache writes are separate categories.

For the Opus cohort, repricing moves the modeled turn from $0.064 to $0.109, or 1.71×. For the Fable cohort, Fable 5.2 prices the same turns at $0.228 each versus $0.284 at Fable 5.1 rates and about $0.123 at Opus rates. At Opus prices, cache reads represent 29.5% of the Opus cohort’s modeled bill and 15.3% of the Fable cohort’s. With every other category priced at 2× and reads at 1×, the premium is simply 2 − the Opus bill share from cache reads.

The cohorts were collected under different task selection. The replay holds token counts fixed and uses rounded published aggregates. It estimates the rate change and the effect of cache mix. Actual Fable 5.2 task cost still depends on its output length, tool calls, retries and how many turns it takes to finish. These were Claude Code usage logs repriced at API list rates, separate from subscription charges. Original study · Published cohort totals.

Where does Fable 5.2 beat Opus 5.5?

Anthropic’s launch results show the largest gaps in this selection on agentic science and the Remote Labor Index. On FrontierCode 2.0, Fable 5.2 at high scored 45.3% versus 45.5% for Opus 5.5 at max, with Anthropic reporting roughly half the cost per task for Fable. Fable reached 49.1% at xhigh. The comparison is specific to that coding harness and those effort settings. The price table alone would predict the opposite ordering on identical tokens. Anthropic’s benchmark and efficiency results.

These are launch results, mostly at each model’s highest available effort. Benchmark denominators and scoring differ. For example, the OSWorld score covers a partial offline set. Anthropic also demonstrated a Lean formalization using roughly 1,000 agents in a single run per model. That illustrates what parallel search can accomplish, with replication and cost per accepted proof still open for buyers. Launch methodology and demonstration · System card.

Knowledge work gives a different ordering. On the Artificial Analysis AA-Briefcase v1.1 preference leaderboard, Fable 5.2 at max scores 1,785 Elo versus Opus 5.5 at 1,800 and Sonnet 5.5 at 1,813, all with default fallbacks. The published intervals for Fable and Opus overlap. Anthropic’s rubric view of the same benchmark shows Fable ahead of Opus, 66.1% to 61.2%. A checklist and a relative preference ranking can favor different outputs. Use whichever scoring method resembles how your team accepts work. AA-Briefcase results · Rubric results.

How much does higher effort buy on AutomationBench?

Zapier’s AutomationBench uses 657 tasks in simulated business systems and checks whether the resulting state satisfies every required assertion. Fable 5.2 at max passed 338 tasks, 23 more than at xhigh. That put it below Gemini 4 Carbon at 53.9%. Zapier’s report distinguishes Opus 5.5 with default fallbacks (42.47%) from the no-fallback comparison (40.0%). Check that configuration before comparing model names across tables. Zapier’s release evaluation.

Max bought 3.5 additional percentage points over xhigh, while estimated spend per pass rose from $4.77 to $9.01. Using the rounded costs per attempt, running all 657 tasks at max instead of xhigh adds roughly $1,544 for 23 net additional passes, or $67 per net additional pass. That measures the difference between two full-run policies. A selective retry policy needs the paired outcomes for tasks that xhigh actually failed.

From low to max, Zapier reports reasoning tokens rising from about 2,100 to 46,700 per task and tool calls rising from 58.4 to 96.9. Marketing peaked at xhigh in the same run. Compare settings by workload and include failed attempts in the spend. Zapier priced observed usage with launch rates and cache assumptions. Its Opus and Sonnet comparison estimates omit cache writes, so direct cross-model cost ratios from that report have uneven accounting. Effort, cost and methodology.

What does Fable need to earn its premium?

Cost per accepted task = total spend across attempts ÷ accepted outcomes. Include token charges, tools, retries, subagents and review time if they change with the route. For one attempt per task, using model costs only, Fable ties Opus when its success rate equals the Opus success rate multiplied by the Fable-to-Opus attempt cost ratio. Using the unrounded replay ratio for the Opus mix as an illustrative attempt-cost proxy gives these thresholds.

At that ratio, above about 58.6% Opus success, Fable can’t win on token spend per accepted task through a higher pass rate alone. It can still win if it uses fewer billable tokens per task, saves enough human review time or handles work whose successful outcome is worth more. Benchmark pass rates belong to their own harnesses. Estimate the threshold using your own per-task spend and acceptance rates.

Measure escalation as a policy. If the first attempt costs c and succeeds with probability p, and Fable costs f and succeeds on fraction q of the first model’s failures, expected spend is c + (1 − p) × f and expected success is p + (1 − p) × q. With a $0.06 first attempt at 50% success and a $0.10 escalation at 90% success on those failures, the policy costs about $0.116 per accepted task, versus $0.120 for one first-model attempt. If conditional rescue is 75%, the same policy costs about $0.126. Measure q on the failures you would actually escalate.

When does Fast mode pay for itself?

Anthropic reports up to 30% faster output than Fable 5.1 in standard mode, and up to 2.5× faster generation when Fast is enabled on 5.2. End-to-end savings depend on the share of elapsed time spent generating tokens. If generation takes 20% of wall time, even 2.5× faster generation saves 12% overall. If generation takes 80%, the same speedup saves 48%. Anthropic’s speed claims · Fast mode pricing.

On the historical Opus token mix, Fast would raise the modeled Fable turn from $0.109 to about $0.219, assuming identical tokens. In the 100-second example, that is about $33 of extra token spend per hour of wall time saved if the savings accrue on the critical path. Compare the premium with measured p95 job time and the value of avoiding a timeout or keeping an operator moving. Tool waits, queue time and parallel branches can absorb most of a generation-speed gain.

What should teams check before migrating?

1. Run matched tasks. Compare your current Opus or Sonnet configuration against Fable at medium, high and xhigh with the same prompts, tools and acceptance criteria. Repeat a subset and inspect paired wins and losses. Compare max where the task justifies the added budget.
2. Capture the full job. Record model, provider, effort, speed, five token buckets, tool and subagent spend, cache writes by TTL, retries, fallbacks, elapsed time and reviewer minutes. Put refusals, timeouts and rejected outputs in the denominator for the attempt-level pass rate. For cost per accepted task, include their spend in the numerator.
3. Exercise conversation replay. Fable 5.2 checks the prefix bound to preserved thinking on every account. Editing earlier messages, system instructions or tool definitions can invalidate those blocks. Switching away from Fable can drop reasoning blocks that the next model cannot read, which can increase latency and output spend after the switch. Test rollback and replay with actual multi-turn fixtures. Anthropic also documents a computer-use tool migration outside Bedrock and a model compatibility check for advisors.
4. Verify capacity and data rules. Fable 5.2 has a separate API rate limit from Fable 5.1 and lacks Priority Tier support. Anthropic requires 30-day retention on the Claude API unless it expressly authorizes zero data retention. Enterprise Frontier Safeguards was in private beta at launch. Review access and retention for the exact account and hosting route used by your workload.

Track whether a reply came from the requested model or from a fallback. Anthropic says its Claude products can use Opus 4.8 for some blocked cyber requests and Opus 5 for some blocked biology requests. API refusal and fallback behavior depends on the route and configuration. Keep those outcomes visible in your quality and cost reports. Migration guide · Refusals and fallback · Launch safeguards and retention.

Does Claude Pro include Fable 5.2?

Anthropic’s launch offer gives Claude Pro and standard Team seats access to Fable models for up to 50% of weekly usage through November 16, 2026 at 11:59 p.m. Pacific time. Anthropic says Fable returns to usage credits on those plans after the promotion or after that included share is used, if credits are enabled. Eligible plans and seats also have temporary reset offers. The API list prices above model token charges. Plan quotas and usage credits follow the subscription terms. Claude launch offer terms.

What is the launch-day routing decision?

If you already pay for Fable 5.1, start by replaying its workloads on 5.2 with the same effort, then test lower settings where completion stays high. If Opus 5.5 already clears your acceptance bar, evaluate Fable on expensive failures such as cross-repo fixes, difficult research or work that repeatedly comes back from review. Promote a route when the paired evaluation shows lower cost per accepted outcome or when the incremental successes justify the incremental spend. This analysis combines published evaluations and a historical rate replay. The rate replay estimates what identical historical usage would cost on Fable 5.2. First-hand Fable 5.2 runs are still needed to measure how its behavior changes that usage.

Related

• Fable 5.1 vs Opus 5.5 on our Claude Code logs
• Cost per task across frontier models
• Opus 5.5 cost by effort level

Want this applied to your stack? Bring the provider bills, gateway logs, and top workflows; we will map the cost drivers and savings path. Book a free audit →

Back to finopsllm.com