Hours after launching Claude Fable 5.2, Anthropic published a practical answer to the question its own lineup raises: which Claude model should you actually use? The claude.dev guide, written by Addy Osmani and dated November 9, recommends a surprisingly conservative default — not the new flagship, but Claude Opus 5.5 at medium effort — and spends most of its length on how to save money by mixing models and tuning effort.
Quick Take
- Default: “Start most work on Opus 5.5 at its default effort level,
medium.” It is the default in Claude Code and costs half as much per token as Fable 5.2. - Move down: Sonnet 5.5 for well-scoped work like bug fixes; Haiku 5.5 for classification, routing, extraction and subagent side tasks.
- Move up: Fable 5.2 for the hardest work in coding, finance and the sciences, when Opus 5.5 at
highstill falls short, or for long unattended runs. - The order of fixes: first let the model verify its own work, then raise effort one level, then move to the next model up.
- The cost trick: run everything at low effort and re-run only failures at high — about 97% pass rate at $0.17 per task in Anthropic’s SWE-bench Pro test, versus 95.3% at $0.29 all on high.
What happened
The @ClaudeDevs account announced the guide on the evening of November 9, saying it “breaks down when to use Fable 5.2, Opus 5.5, Sonnet or Haiku” and covers “when to adjust effort instead of switching models, and how to save cost by putting subagents on smaller models.” The post passed 10,000 likes, and developers quickly turned it into checklists and Claude Code configurations. The same advice appears in Anthropic’s Fable 5.2 launch post, which says that “for most coding and knowledge tasks, we recommend that users start with Claude Opus 5.5, our strongest daily driver,” and in the platform docs.
Background: four models, five effort levels
Anthropic’s current generally available lineup has four models, all with 1M-token context windows and 128K-token maximum output. According to the guide, list prices per million input/output tokens are $0.10/$0.50 for Haiku 5.5 (prompts up to 100K tokens; $0.50/$2.50 above), $2/$10 for Sonnet 5.5, $4/$20 for Opus 5.5 and $8/$40 for Fable 5.2. Claude Mythos 5.2 is available only through Anthropic’s verification programs. Each model also has an “effort” parameter that trades capability for cost and latency. Default effort differs by model and surface: Opus 5.5 and Haiku 5.5 default to medium on the API, while Fable 5.2 and Sonnet 5.5 default to high (medium in Claude Code).
Context: Opus 5.5 recently drew complaints about a behavioral shift in early November, covered in Claude Opus 5.5 nerf complaints and the data, and i10x reported earlier on Haiku 5.5 as a subagent model.
The guide’s task table
| Task | Starting model | Starting effort |
|---|---|---|
| Everyday engineering, debugging, multi-PR features | Opus 5.5 | medium, then high |
| Classification, routing, tagging, high-volume extraction | Haiku 5.5 | medium |
| Subagents that search files, read logs or summarize | Haiku 5.5 or Sonnet 5.5 | model default |
| Chat and interactive sessions | Sonnet 5.5 (Haiku 5.5 for simple chat) | low or medium |
| Well-scoped coding: bug fixes, spec-driven work | Sonnet 5.5 | medium, then high |
| Computer use and vision-heavy workflows | Opus 5.5 | medium |
| Major refactors, migrations, leading many subagents | Fable 5.2 | medium, then high |
| Deep research, scientific or quantitative analysis | Fable 5.2 | medium, then high |
| Not sure yet | Opus 5.5 | medium |
How the advice works in practice
Effort before model. Effort is “a finer dial for both cost and capability,” the guide says, but “past high, one more effort level can cost as much as a bigger model.” Its SWE-bench Pro numbers for Opus 5.5: medium scored about 2.5 points lower than high for about 70% of the cost; low about 8 points lower for a third of the cost; xhigh about 1.4 points higher for 2.5 times the cost. Higher effort “won’t give a model knowledge or skill it doesn’t have.”
Price per task, not per token. On Anthropic’s Terminal-Bench 3.0 runs, Fable 5.1 at max scored 58.0% while Opus 5.5 at high scored 58.9% “with half the tokens.” On FrontierCode 2.0, Fable 5.2 at high roughly matched Opus 5.5 at max for about half the cost — so “before you raise Opus 5.5 past high on a hard task, test Fable 5.2.” On one 20-problem benchmark, two problems carried 43% of the spend.
Subagents on smaller models. In Claude Code, a subagent definition can set model: haiku or model: sonnet, and CLAUDE_CODE_SUBAGENT_MODEL sets the default. The guide recommends small models only for work “where a mistake is cheap to spot, like finding files, running tests and reading logs,” while code edits and judgment calls stay on the main model. One notable change: the built-in Explore subagent now runs on the session’s model instead of always on Haiku, so the guide suggests defining a custom Explore subagent with model: haiku.
Switching costs. The prompt cache belongs to the model that wrote it, so switching models mid-conversation re-pays cache writes. Thinking blocks are only partly readable across models — “a move from Fable 5.2 to any other generally available model drops its reasoning.”
Advisors and orchestration. The guide is cautious on its advisor tool (beta): in one internal pairing, an Opus 5.5 executor with a Fable 5.1 advisor scored 1.7 points higher “for about 2.1 times the money.” Bulk orchestration did better: on a defect-search benchmark over about 22 million tokens of code, an Opus 5.5 lead with Sonnet 5.5 workers cost about a third less than Opus 5.5 alone and scored about the same.
Pros and cons
For users
- Pro: Clear defaults reduce guesswork, and the advice helps Pro and Max subscribers stretch usage limits by not running everything on the biggest model.
- Con: The recommendations assume Claude Code and API controls that casual chat users don’t see.
For developers
- Pro: Concrete numbers, config snippets and a compatibility checklist (for example, forced
tool_choicereturns a 400 error on Sonnet 5.5, Opus 5.5 and Fable 5.2) make migrations safer. - Con: Many figures come from Anthropic’s internal runs, and the guide itself warns that multi-model setups must beat every single-model point on your own cost-vs-score chart.
For companies
- Pro: Low-effort-first with high-effort retries and cheaper worker models offer measurable savings for large agent fleets.
- Con: Routing adds operational complexity — cache misses on switches, dropped reasoning and different API quirks per model.
What X is saying
Which Claude model should you build with? Our new guide breaks down when to use Fable 5.2, Opus 5.5, Sonnet or Haiku. It also covers when to adjust effort instead of switching models, and how to save cost by putting subagents on smaller models.
— ClaudeDevs (@ClaudeDevs) November 9, 2026 · View on X
before you blame fable 5.2 for eating your limit, here’s what anthropic says each claude model is for 👀
— Aditya Umale (@AdityaUmale17) November 10, 2026 · View on X
Claude Code’s built-in Explore subagent used to always run on Haiku. Now it runs on your main model. If that’s Opus 5.5, or you just switched to Fable 5.2, Explore is burning big-model quota for nothing every time it reads files or searches code. So define your own subagent named Explore and set model: haiku.
— Vox (@Voxyz_ai) November 10, 2026 · View on X
@AdityaUmale17 distilled the guide’s “cheapest trick”: “run everything on low effort, rerun only the failures on high.” @Voxyz_ai shared a long prompt asking Fable 5.2 to audit a Claude Code setup against the guide and to “Show me the recommendations first. Don’t change anything yet.” Other posts circulated more elaborate setups pairing Sonnet sessions with Haiku subagents and an Opus advisor; i10x could not verify every command-line flag in those posts against Anthropic’s documentation.
What Grok says
Asked to summarize the guide, Grok answered (in Chinese; i10x translation): default to Opus 5.5 at medium effort, “the best value, first choice for everyday complex coding and knowledge work”; Haiku 5.5 for high-concurrency simple tasks such as classification, extraction and subagents; Sonnet 5.5 for well-scoped daily tasks at about half Opus’s cost; and Fable 5.2 for the hardest technical projects, deep research and large refactors.
Critical analysis
It is unusual for a vendor to launch its most expensive model and, the same day, tell most customers not to start with it. That candor is the guide’s strength — and partly self-interest: routing users away from Fable also eases load during the launch promotion and keeps heavy users from hitting limits. The core methodology, however, is sound and vendor-neutral: verification first, effort second, model third; measure cost per finished task; keep judgment on the strongest model.
The weak point is evidence. The guide’s numbers are mostly Anthropic’s own runs on Anthropic’s own benchmarks or internal tasks, and several (the advisor results) are explicitly near “run-to-run noise.” It also quietly documents friction that makes multi-model setups harder than the X threads suggest: cache invalidation, dropped reasoning across tiers, and different 400-error rules per model.
What’s still open
- Will Anthropic publish the underlying evaluation data for the cost-per-task comparisons?
- Will Claude Code restore Haiku as the default for the Explore subagent, or keep it tied to the session model?
- How will the advisor tool perform once it leaves beta?
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated illustration. Read our AI Use Policy. Spotted an error? Request a correction.
This independent, research-based i10x analysis draws on the claude.dev guide, Anthropic’s platform documentation and Fable 5.2 launch post, and public posts on X through November 10, 2026. Related: Claude Fable 5.2 and Mythos 5.2 launch and Anthropic’s single-agent loop and subagents guidance.
i10x Perspective
Model choice is becoming an engineering discipline rather than a brand preference. Anthropic’s guide is essentially a playbook for cost-aware agent design: a capable default, cheap workers for checkable tasks, and the top model reserved for work that genuinely needs it. Teams that build evaluation harnesses for their own tasks — with outcome checks and per-task cost tracking — will get far more from this generation than teams that simply switch to the newest model and hope.
