Which Claude model to use (guide of 9 November 2026)
Definition
Anthropic’s own answer to “which model for which task”, published on claude.dev by Addy Osmani. Four generally available models: Haiku 5.5, Sonnet 5.5, Opus 5.5 and Fable 5.2 (Mythos 5.2 is verification-programme only). Every number below is from Anthropic’s runs, some on internal benchmarks, none reproduced outside (reported). The guide itself says to test candidates on your own tasks.
Rule
- Start on Opus 5.5 at its default effort, medium. It is the Claude Code default and half the price of Fable 5.2.
- When a result is not good enough, three steps in this order: make sure the model can verify its own work (test, build, script that hits the endpoint); raise effort one level; move up a model, at a lower effort than you used on the smaller one.
- Down to Sonnet 5.5 for scoped work (bug fix, spec-driven feature, repeated investigation or review). A clear spec does not guarantee Sonnet: a migration or a performance job can be specified and still need Opus.
- Down to Haiku 5.5 for classification, routing, extraction and subagents that search files, read logs or summarise.
- Up to Fable 5.2 for major refactors and migrations, architecture, leading many
subagents, deep research, long unattended runs, or when Opus at
highstill falls short. Test Fable before raising Opus pasthigh. - Subagents on small models, edits on the main one. Keep the expensive model’s context for decisions. Small models only where a wrong answer is cheap to spot.
- Switch models at a natural break. The prompt cache belongs to the model that wrote
it, so the first turn on the new model pays the cache-write price for the whole
conversation. Run
/compactor start fresh with a written plan first. - Keep a multi-model setup only if it beats every single-model point on a score-versus-spend chart from your own tasks.
Numbers
Prices per million tokens, Claude API list:
| Haiku 5.5 | Sonnet 5.5 | Opus 5.5 | Fable 5.2 | |
|---|---|---|---|---|
| Input / output | $0.10 / $0.50 (prompt <=100K); $0.50 / $2.50 above | $2 / $10 | $4 / $20 | $8 / $40 |
| Cache read | $0.01 / $0.05 | $0.10 | $0.20 | $0.20 |
| Default effort (API) | medium | high (medium in Claude Code) | medium | high (medium in Claude Code) |
| Fast mode | no | no | yes | yes |
All four have a 1M context window and 128K max output. Haiku 5.5 is the only model whose price depends on prompt length; over 100K it is still a quarter of Sonnet’s per-token price, and under 100K its input is a twentieth of Sonnet’s.
Effort on SWE-bench Pro, Opus 5.5, relative to high:
| Effort | Score | Cost |
|---|---|---|
| low | about -8 points | about 1/3 |
| medium | about -2.5 points | about 70% |
| xhigh | about +1.4 points | 2.5x |
Other Anthropic runs:
- Low effort for everything, re-run only failures at
high: about 97% passed at about $0.17 a task, against 95.3% at $0.29 with all tasks athigh. Needs an automatic check. - Terminal-Bench 3.0: Fable 5.1 at
max58.0%, Opus 5.5 athigh58.9% with half the tokens. - FrontierCode 2.0: Fable 5.2 at
high45.3% vs Opus 5.5 atmax45.5%, about half the cost per task. - 20-problem search benchmark: two problems carried 43% of the spend. Price the hard tasks, not the typical ones.
- Opus 5.5 lead with up to 25 Sonnet 5.5 workers over about 22M tokens of source: about a third cheaper than Opus 5.5 alone, about the same score.
- Advisor tool, Opus 5.5 executor at
highwith a Fable 5.1 advisor: 90.1% at $2.92, +1.7 points for about 2.1x the money, at the edge of noise. - Cache read at 150K context costs about $0.03 a turn on Opus 5.5, about $0.004 at 20K.
Claude Code specifics that bite
- Built-in Explore no longer always runs on Haiku. In current versions it inherits the
session model (capped at Opus on the API). To keep exploration cheap, define your own
subagent named
Explorewithmodel: haiku. haikualias resolves to Haiku 5.5 on the Anthropic API only, and needs Claude Code 2.1.293 or later. On Bedrock, Google Cloud, Foundry and Claude Platform on AWS it still means Haiku 4.5.ANTHROPIC_DEFAULT_HAIKU_MODELpins it.model: haiku|sonnetin a subagent definition, orCLAUDE_CODE_SUBAGENT_MODELas the default for subagents that name no model.opusplanputs code edits on Sonnet, the reverse of rule 6. Measure before making it a default.- A move down a tier drops the earlier thinking blocks (a model never reads blocks from a model above it); Fable 5.2 to any other model drops them too.
- Flagged cybersecurity requests can silently re-run on an older model (Opus 4.8, biology on Opus 5); the transcript says so.
Exceptions
- Small models are weaker in long agentic loops; they fit high-volume work with checkable output. Computer use and vision-heavy loops start on Opus 5.5.
- For replies a person waits on, Sonnet 5.5 at
lowormedium. - Effort adds latency and gives no knowledge the model lacks; for that, move up a model.
- Own API code:
tool_choiceany/toolreturns 400 on Sonnet 5.5, Opus 5.5 and Fable 5.2 (useautowithstrict: truetools). Fable 5.2 returns 400 if anything before one of its thinking blocks changed, so history must be append-only. Haiku 4.5 to 5.5: replacebudget_tokenswith effort, drop non-defaulttemperature/top_p/top_k, endmessageson a user turn. Build requests for the strictest model and routing becomes config; let code, not a model, pick the model when the rule is simple.
How it reads against our setup
- Delegation: this is lead-model-delegation-economics from the vendor side. Anthropic says the same thing about the boundary as cheap-model-delegation-boundary: cheap models read, the lead edits and judges. Only the Opus-lead-plus-Sonnet-workers figure is a cost number for delegation, and it is on a bulk-reading task.
- Effort: confirms claude-5-5-effort-and-cost (effort is the finer dial) and gives
its missing curve. Our global
highsits one step above Anthropic’s start, justified only by the Stop-hook gate;xhighbuys 1.4 points for 2.5x. - Haiku: the September note said “recheck when Haiku 5.5 ships”. It shipped. The
override
ANTHROPIC_DEFAULT_HAIKU_MODEL=claude-sonnet-5-5costs 20x Haiku’s input price for every lookup subagent that nameshaiku. - Verification before escalation (rule 2, first step) is the sensor argument of harness-cost-by-model-tier.
Learned from
Read 10 Nov 2026 after a Telegram digest item. Primary page fetched in full; the date and author came from a secondary summary (the page shows neither).