Start with Claude Opus 5.5 at medium effort and move to Claude Fable 5.2 only when a measured task needs it. That is Anthropic's own advice: its Fable 5.2 launch post says "For most coding and knowledge tasks, we recommend that users start with Claude Opus 5.5, our strongest daily driver". This guide is accurate as of 11 November 2026, and Fable 5.2 is two days old, so treat every figure as an early reading. It turns cost per finished task into rupees at the RBI reference rate of Rs 93.8579 per US dollar on 9 November 2026, rounded to the nearest rupee, and ends with a routing plan a small team can copy.
Which Claude model should you use first?
Make Opus 5.5 your default for coding and knowledge work. Fable 5.2 costs exactly twice as much per token, and it cost about 1.6 times as much per run in a 52-run independent test on a production codebase. It earns that premium on long, open-ended jobs and on reviews of risky changes. Sonnet 5.5 and Haiku 5.5 take the work below that line.
Why is list price the wrong comparison?
Anthropic's cost guide says it directly: "Compare models on cost per completed task, not per token, because you pay for completed tasks." Token prices set the rate. Effort, retries and thinking decide the bill.
- •Fable 5.2: about Rs 751 input and Rs 3,754 output per million tokens ($8 and $40).
- •Opus 5.5: about Rs 375 input and Rs 1,877 output ($4 and $20).
- •Sonnet 5.5: about Rs 188 input and Rs 939 output ($2 and $10).
- •Haiku 5.5: about Rs 9.40 input and Rs 47 output ($0.10 and $0.50) for prompts up to 100,000 tokens. A request over that line pays Rs 47 and Rs 235 ($0.50 and $2.50) for the whole request.
- •Cache reads cost the same on Fable 5.2 and Opus 5.5, about Rs 19 per million ($0.20), so in cache-heavy loops the output price drives the gap.
- •Batch is 50% off on all four. Fast mode, listed for Fable 5.2 and Opus 5.5 of these four, doubles the token price.
What does one task cost at each effort level?
Artificial Analysis is an independent evaluator. Its cost per task covers its own index tasks and the leaderboard page does not define the calculation, so use it for the shape of the curve and not as a forecast. On this index, Fable 5.2 at high (62 for Rs 852) beats Opus 5.5 at max (59 for Rs 942) on score and on money. Fable 5.2 at low (54 for Rs 291) also beats Opus 5.5 at high (52 for Rs 325). Max on Fable 5.2 buys three points over high for nearly three times the money.
| Model and effort | AA index score | USD per task | Rupees per task | Minutes per task |
|---|---|---|---|---|
| Fable 5.2, max | 65 | $26.54 | Rs 2,491 | 44.4 |
| Fable 5.2, xhigh | 62 | $14.83 | Rs 1,392 | 26.1 |
| Fable 5.2, high | 62 | $9.08 | Rs 852 | 17.5 |
| Fable 5.2, medium | 56 | $3.80 | Rs 357 | 7.9 |
| Fable 5.2, low | 54 | $3.10 | Rs 291 | 6.4 |
| Opus 5.5, max | 59 | $10.04 | Rs 942 | 24.7 |
| Opus 5.5, xhigh | 57 | $6.05 | Rs 568 | 15.9 |
| Opus 5.5, high | 52 | $3.46 | Rs 325 | 9.6 |
| Opus 5.5, medium | 48 | $2.45 | Rs 230 | 7.1 |
| Opus 5.5, low | 35 | $0.98 | Rs 92 | 2.9 |
| Sonnet 5.5, high | 42 | $1.64 | Rs 154 | 6.9 |
| Haiku 5.5, max | 36 | $0.41 | Rs 38 | 14.8 |
| Haiku 5.5, medium | 25 | $0.09 | Rs 8 | 3.9 |
What do real coding tests say about cost per finished task?
On coding work the picture shifts back towards Opus 5.5. At medium effort, two separate sources put Fable 5.2 at about 1.6 times the cost. Every row names who ran the test. None of this is a production bill: we found no named company that has published one comparing these two models.
| Test and who ran it | First option | Second option | What to read from it |
|---|---|---|---|
| 52 runs each, TypeScript monorepo of about 3,700 files, medium effort. Independent, paddo.dev | Fable 5.2: Rs 231 per run ($2.46), 40 of 52 passed | Opus 5.5: Rs 142 per run ($1.51), 44 of 52 passed | 1.6x. The author judged all five of Fable's failed bug runs to be correct fixes an exact-match test rejected. One codebase. |
| Internal agentic-coding benchmark. Vendor-reported, Anthropic | Fable 5.2 medium: Rs 126 per attempt ($1.34), 88.9%, single run | Opus 5.5 medium: Rs 79 per attempt ($0.84), 86.6% | 1.6x again, for 2.3 points. |
| SWE-bench Pro subset, 463 problems. Vendor-reported, Anthropic | Fable 5.2 low: Rs 37 per solved task ($0.39), 94.6% | Opus 5.5 medium: Rs 21 ($0.22), 93.1%. High: Rs 28 ($0.30), 95.5% | Opus 5.5 is cheaper per solved task at a similar pass rate. |
| FrontierCode 2.0. Run by Cognition, published in Anthropic's system card | Fable 5.2 high: Rs 631 ($6.72), 45.3%. Xhigh: Rs 1,332 ($14.19), 49.1% | Opus 5.5 max: Rs 1,203 ($12.82), 45.5% | The half-the-cost claim compares Fable at high with Opus at its dearest setting. |
| Opus 5.5 at low, failures re-run at high, SWE-bench Pro subset. Vendor-reported, Anthropic | Low then retry: Rs 16 per task ($0.17), about 97% passed | Every task at high: Rs 27 ($0.29), 95.3% passed | Needs a check you can run automatically. |
| Agent team moved from mostly Opus to a three-tier mix. Self-reported, Karun Japhet | After: about Rs 188 per story ($2) | Before: about Rs 404 per story ($4.30) | One day of after-data, versions not named, and context trimming landed together, so the saving cannot be split. |
When does Fable 5.2 earn its price?
Anthropic's own trigger: "Use Claude Fable 5.2 for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5.5 at higher effort still fall short." We would advise paying double in these cases.
- •Long, open-ended jobs nobody can check step by step. Anthropic's system card reports that Fable 5.2 "completed 41.3% of freelance projects at least as well as a human professional, compared with 21.3% for Opus 5.5" on the Remote Labor Index. Vendor-reported.
- •Reviews of changes touching money, routing or another system. In the paddo.dev test Fable 5.2 flagged a routing gap in Opus 5.5's passing fix in 4 of 4 reviews, and Opus 5.5 in 1 of 4. That is one bug in one codebase, at about Rs 113 per Fable review ($41.01 over 34).
- •Hard coding where Opus 5.5 has topped out. On FrontierCode 2.0, Fable 5.2 at xhigh reached 49.1% against Opus 5.5's best of 45.5%, for about 11% more per task.
- •Tasks where Opus 5.5 at high has already failed your own checks.
When does Opus 5.5 win?
Opus 5.5 wins on routine work and on two knowledge-work boards, and some of this evidence comes from Anthropic itself.
- •Routine tickets: the paddo.dev author recommends Opus 5.5 here, with the same test results at about 60% of the cost.
- •Deep research: on 44 DeepResearch Bench II tasks at default effort, Fable 5.2 cost 2.9 times as much as Opus 5.5 and scored 1.9 points lower. Vendor-reported.
- •Knowledge-work Elo: GDPval-AA v2.1 has Opus 5.5 at 1866 and Fable 5.2 at 1831. AA-Briefcase v1.1 has 1800 and 1785. These are the two results in the system card's table where Fable trails.
- •Code tasks with images: Vals AI's Fable 5.2 page shows SWE-bench Multimodal at 62.29%, with Opus 5.5 at 70.21%. Independent.
- •Data retention: Fable 5.2 requires 30-day retention and is unavailable under zero data retention unless Anthropic expressly authorises it. See /blog/fable-5-2-30-day-retention-inference-hooks-dpdp-claude-code-security.
Where do Sonnet 5.5 and Haiku 5.5 fit?
Anthropic's guidance is to "Move down to Sonnet 5.5 for well-scoped everyday work, like a bug fix or a feature built against a spec", and to use Haiku 5.5 for the simplest high-volume work such as classification, routing and extraction. A clear spec alone does not make a task easy: Anthropic warns that a migration can still warrant Opus 5.5.
- •Sonnet 5.5 at its default high effort and Opus 5.5 at low are close to a tie on Anthropic's SWE-bench Pro subset: 87.6% for Rs 10 per solved task ($0.11) against 87.4% for Rs 11 ($0.12). Use Sonnet 5.5 at low or medium when a person is waiting.
- •Opus 5.5 leading Sonnet 5.5 workers cost about a third less than Opus 5.5 alone with about the same score, on work larger than one context window. Putting Fable 5.2 in the lead cost 19% more with about the same score. Vendor-reported.
- •Haiku 5.5 scored 85% on GPQA Diamond against 91% for Opus 5.5, at about a twentieth of the cost per correct answer. Vendor-reported.
- •Haiku 5.5 as a coding subagent under Opus 5.5 cut cost 32% and pass rate about 4 points, so try lower effort first. Vendor-reported.
A one-page routing plan for a small team
Take a team closing 400 tickets a month. This is a planning model, not a measured bill, so replace each rate after a month of your own logs. Opus 5.5 and Fable 5.2 ticket rates are the paddo.dev per-run costs. Sonnet 5.5 at $0.76 is our assumption, half of Opus 5.5, because Anthropic's subset shows $0.11 against $0.22 per solved task. Haiku 5.5 uses Anthropic's $0.011 per passed attempt on a 79-task set. The long-job rate is FrontierCode 2.0 at high.
- •Arithmetic: $4.40 + $167.20 + $211.40 + $95.20 + $33.60 = $511.80, multiplied by 93.8579, is about Rs 48,036.
- •All 400 tickets on Opus 5.5 at medium: 400 x $1.51 = $604, about Rs 56,690. The routed plan is 15% lower and still buys Fable 5.2 on the 40 tickets that need it.
- •All 400 on Fable 5.2 at medium: 400 x $2.46 = $984, about Rs 92,356. The routed plan is 48% lower.
| Tier and work | Model and effort | Tasks | USD each | Rupees a month |
|---|---|---|---|---|
| Triage: classify and route every ticket | Haiku 5.5, medium | 400 | $0.011 | Rs 413 |
| Well-scoped: bug fix or feature against a spec | Sonnet 5.5, high | 220 | $0.76 | Rs 15,693 |
| Default: everything else | Opus 5.5, medium | 140 | $1.51 | Rs 19,842 |
| Risky: money, routing, integrations | Opus 5.5 builds, Fable 5.2 reviews | 35 | $2.72 | Rs 8,935 |
| Long job: hours of agent time | Fable 5.2, high | 5 | $6.72 | Rs 3,154 |
| Total | 400 tickets | $511.80 | Rs 48,036 |
Set effort on every request
Fable 5.2's API default is high and Opus 5.5's is medium, so a request that omits effort runs one level higher on Fable 5.2. In Claude apps and Claude Code the Fable 5.2 default is medium.
Give each task a check
Anthropic's first step when a model falls short is to let it verify its own work with a test, a build or a script.
Escalate in order
Raise effort one level, then move up a model at a lower effort than you used on the smaller one.
Re-run failures instead of raising everything
Anthropic's table puts this at about 40% less than running every task at high, at the same pass rate or slightly better.
Log rupees per finished task by tier
Anthropic's advice is to "Recheck at every model release, because releases move both the capability gap and the price ratio."
How do you run low effort first and retry failures?
The pattern needs a check you can run automatically. In this Python sketch, the task object and its check method stand in for your own prompt and test. Set max_tokens generously, because thinking counts against it.
What mistakes inflate a Fable 5.2 bill?
Most overspend comes from settings, and the per-token price is rarely to blame.
⚠️Reaching for max effort
Consequence: On FrontierCode 2.0, Fable 5.2 at max scored 45.7% for Rs 4,195 ($44.70), while xhigh scored 49.1% for Rs 1,332 ($14.19).
Solution: Stop at xhigh, and use that only where you have measured a gain.
⚠️Leaving Fable 5.2 at xhigh in an agentic tool without watching subagents
Consequence: Anthropic observed it starting subagents to review its own work at higher effort, and warns that "Subagents can account for a large part of what a session costs".
Solution: Cap effort per task class and read cost per session, not per message.
⚠️Assuming Claude Pro keeps including Fable
Consequence: Pro inclusion is a one-week promotion that ends on 16 November 2026 at 11:59 pm PT. After that, Fable runs on usage credits.
Solution: Plan around credits, or see /blog/claude-pro-max-ultra-usage-limits-november-2026-india for which plan fits.
How Tech Arion helps with this decision
Tech Arion is a Hyderabad software and AI consultancy that builds with Claude Code. We help teams choose a default model and report spend as rupees per finished task. Moving API code? Read /blog/claude-api-2026-10-21-fable-5-2-migration-breaking-changes first.
Frequently asked questions about Claude Fable 5.2 vs Opus 5.5
Short answers, as of 11 November 2026.
Frequently Asked Questions
Find out what your Claude work costs per finished task
Tech Arion can baseline your current spend in rupees, build the checks that make routing safe, and set up the Opus 5.5, Fable 5.2, Sonnet 5.5 and Haiku 5.5 split for your workload.
Sources & References
Sources opened on 11 November 2026. Benchmarks are labelled in the article as vendor-reported or independent.
- 1.
Anthropic. (9 Nov 2026). Claude Fable 5.2 and Claude Mythos 5.2 launch post: recommendation to start with Opus 5.5, $8/$40 pricing, medium as the default effort in Claude apps and Claude Code.
View Source - 2.
Anthropic. (2026). Migrating to Claude Fable 5.2: default effort high against medium on Opus 5.5, self-review subagents at higher effort, 30-day retention.
View Source - 3.
Anthropic. (2026). Optimizing for cost and intelligence: cost per completed task, SWE-bench Pro subset, DeepResearch Bench II, orchestrator and Haiku subagent results, GPQA Diamond, re-running failures.
View Source - 4.
Anthropic. (2026). Claude platform pricing: Fable 5.2, Opus 5.5, Sonnet 5.5 and Haiku 5.5 rates, cache reads, Batch, fast mode, the Haiku 100,000-token step.
View Source - 5.
Osmani, A., Anthropic. (9 Nov 2026). Which Claude model should you build with? Start on Opus 5.5 at medium, escalation order, low-then-retry at $0.17 against $0.29, Sonnet and Haiku guidance.
View Source - 6.
Anthropic. (9 Nov 2026). Claude Fable 5.2 and Claude Mythos 5.2 system card: FrontierCode 2.0 scores and cost per task as run by Cognition (p. 209), Remote Labor Index, GDPval-AA and AA-Briefcase in Table 8.1.A.
View Source - 7.
Artificial Analysis. (Live, 11 Nov 2026). Models leaderboard: Intelligence Index score, cost per task and time per task by effort level. Independent.
View Source - 8.
Vals AI. (9 Nov 2026). Claude Fable 5.2 model page: SWE-bench Multimodal 62.29% with Opus 5.5 at 70.21%. Independent.
View Source - 9.
paddo.dev. (10 Nov 2026). Fable 5.2 vs Opus 5.5: 52 runs per model at medium effort on a production TypeScript monorepo, $127.93 against $78.44, review results. Independent, one codebase.
View Source - 10.
Japhet, K. (10 Nov 2026). Optimising cost for agent teams: cost per story from about $4.30 to about $2, with the author's caveats. Model versions not named.
View Source - 11.
Metropolitan Stock Exchange of India. (9 Nov 2026). RBI reference rate archive: USD/INR 93.8579, used for every rupee figure here.
View Source - 12.
Anthropic Help Center. (Nov 2026). Claude Fable 5.2 launch offers: Pro and standard Team seats included from 9 to 16 November 2026 at 11:59 pm PT, up to 50% of weekly limits.
View Source