Verified against Anthropic's documentation on September 30, 2026.
Picking a Claude model used to be one decision on one axis. Smart, balanced, or fast.
That framing is now half the decision. There is a second control called effort, and Anthropic's own model guide now says tuning effort is often a better lever than switching models.
So you make one choice out of two, then wonder why the bill or the latency looks wrong.
Here are both decisions on one page.
Question one: which model
Four current models. Prices are per million tokens.
Claude Fable 5.1 · In / Out: $10 / $50 · Context: 1M · Default effort (API):
high· Pick it for: Demanding reasoning and long-running agent work, when capability beats costClaude Opus 5.5 · In / Out: $4 / $20 · Context: 1M · Default effort (API):
medium· Pick it for: Long-running agentic coding and knowledge work. The everyday defaultClaude Sonnet 5.5 · In / Out: $2 / $10 · Context: 1M · Default effort (API):
high· Pick it for: Best combination of speed and intelligence. Everyday coding, analysis, high volumeClaude Haiku 4.5 · In / Out: $1 / $5 · Context: 200k · Default effort (API): None · Pick it for: Fastest and cheapest. Real-time, high-volume, and subagent work. No effort control
Still available as legacy models: Fable 5, Opus 5, Opus 4.8, 4.7, 4.6 and 4.5, Sonnet 5 and Sonnet 4.6. Sonnet 4.5 is deprecated and retires November 30, 2026. Claude Mythos 5.1 also exists, by invitation only, through Anthropic's Project Glasswing.
Anthropic's own starting advice: begin with Opus 5.5 for most workloads, and use Fable 5.1 for demanding reasoning and long-horizon agent work, or when your own tests on Opus 5.5 at higher effort still fall short.
The one fact that decides most of these: Opus 5.5 costs 40 percent of Fable 5.1's price per token, and Anthropic says it performs at Fable 5.1's level on most work. Unless you have measured Fable 5.1 winning on your own work, Opus 5.5 is the better default.
The tiebreaker nobody checks: knowledge cutoffs. Fable 5.1, Opus 5.5 and Sonnet 5.5 are all reliable through June 2026. Haiku 4.5 is reliable through February 2025, sixteen months earlier. Haiku 4.5 is also the oldest model still current, and Anthropic's commitment to keep it available runs to at least October 15, 2026, the shortest runway in the lineup. No retirement has been announced, and Anthropic gives at least 60 days' notice before one. Check the deprecations page before you build anything long-lived on it.
By job
Agentic coding, multi-file, long sessions · Model: Opus 5.5
An agent that runs for hours unattended · Model: Fable 5.1
Drafting, analysis, research, everyday chat · Model: Sonnet 5.5
Classification, extraction, tagging at volume · Model: Haiku 4.5
Anything about recent events · Model: Any model, but supply the sources yourself and verify the dates
Subagents inside a larger run · Model: Sonnet 5.5 or Haiku 4.5
You genuinely cannot get a good result · Model: Fable 5.1, then stop
Question two: which effort
Effort controls how many tokens Claude spends on a response. It is a real dial with five settings, and it applies to everything: the answer, the explanations, the tool calls, and the thinking.
low· Cost and speed: Cheapest, fastest · Use it for: Simple tasks, lookups, classification, subagentsmedium· Cost and speed: Balanced. The default on Opus 5.5 · Use it for: Everyday agent work where speed mattershigh· Cost and speed: The default on every other model with effort · Use it for: Complex reasoning, hard coding, real agent workxhigh· Cost and speed: Expensive, slow · Use it for: Long-horizon runs over 30 minutesmax· Cost and speed: No constraint on reasoning spend · Use it for: Frontier problems only
Setting effort to a model's default is identical to setting nothing at all. On Opus 5.5 that default is medium, not high.
Which models take which levels
Not every level exists on every model. This is where instructions copied between models fail.
Claude Fable 5.1 · low / medium / high: Yes · xhigh: Yes · max: Yes
Claude Opus 5.5 · low / medium / high: Yes · xhigh: Yes · max: Yes
Claude Sonnet 5.5 · low / medium / high: Yes · xhigh: Yes · max: Yes
Claude Fable 5 · low / medium / high: Yes · xhigh: Yes · max: Yes
Claude Opus 5 · low / medium / high: Yes · xhigh: Yes · max: Yes
Claude Sonnet 5 · low / medium / high: Yes · xhigh: Yes · max: Yes
Claude Opus 4.8 · low / medium / high: Yes · xhigh: Yes · max: Yes
Claude Opus 4.7 · low / medium / high: Yes · xhigh: Yes · max: Yes
Claude Opus 4.6 · low / medium / high: Yes · xhigh: No · max: Yes
Claude Sonnet 4.6 · low / medium / high: Yes · xhigh: No · max: Yes
Claude Opus 4.5 · low / medium / high: Yes · xhigh: No · max: No
Claude Sonnet 4.5 · low / medium / high: Not supported · xhigh: No · max: No
Claude Haiku 4.5 · low / medium / high: Not supported · xhigh: No · max: No
Haiku 4.5 has no effort control at all. It is already the efficient option.
The per-model starting points
These come from Anthropic's effort documentation, and they differ by model on purpose.
Claude Opus 5.5. Starts at medium, the default. Thinking is always on and cannot be switched off, so effort is the main control over how much it reasons and what a request costs. Anthropic's advice: run your own effort test rather than carrying settings over from an earlier model, and give it a large output limit at the higher levels.
Claude Sonnet 5.5. Defaults to high on the API. Its levels were recalibrated, so the same level does not buy the same amount of thinking it did on Sonnet 5. Start at high unless the work is agentic or latency-sensitive. For agentic coding and multistep tool use, start at medium for well-specified tasks and move to high for harder or longer ones. For chat, start at medium or low. Use xhigh or max only where your testing shows a gain.
Claude Fable 5.1. Start at high, the default. Step up to xhigh or max for the most capability-sensitive agent and coding work. Step down to medium or low for routine or latency-sensitive work once your testing shows quality holds.
Claude Fable 5, now legacy. Effort is the primary control on this model. Start at high. Lower settings on Fable 5 still perform well and often beat earlier models at xhigh.
Claude Opus 5, now legacy. Start at high, the default. Step up to xhigh for demanding coding and agent work, or max when the task justifies unconstrained reasoning spend. Use low and medium freely wherever your testing shows quality holds.
Claude Opus 4.7 and 4.8. Different advice again. Start at xhigh for coding and agent work, and treat high as the floor for anything intelligence-sensitive.
That last line is the trap. The right starting effort on Opus 4.8 is the wrong starting effort on Opus 5.5. And if you moved from Opus 5 to Opus 5.5 without setting effort, your requests now run one level lower than before, because the default dropped from high to medium. Anthropic says it plainly: run a fresh effort test rather than reusing old settings.
The combinations worth memorizing
Everyday coding in Claude Code · Model: Opus 5.5 · Effort:
medium, the defaultA hard refactor across many files · Model: Opus 5.5 · Effort:
high, orxhighif it runs past 30 minutesAn agent running unattended for hours · Model: Fable 5.1 · Effort:
highDrafting, research, summarizing · Model: Sonnet 5.5 · Effort:
highHigh-volume chat or support replies · Model: Sonnet 5.5 · Effort:
lowormediumClassifying or tagging thousands of items · Model: Haiku 4.5 · Effort: n/a
Subagents inside a bigger run · Model: Sonnet 5.5 · Effort:
lowfor simple lookups,mediumfor subagents that use toolsA problem two models already failed · Model: Fable 5.1 · Effort:
max
Five things that will cost you
1. Reusing effort settings across models. The single most likely source of a surprise bill or a surprise latency spike. Opus 4.8 wants xhigh for coding. Opus 5.5 starts at medium. Sonnet 5.5 recalibrated its levels, so high there is not high on Sonnet 5. Same task, different answer.
2. Turning effort down to get shorter answers. Anthropic says it directly for Opus 5: effort controls thinking volume, not visible response length. Effort is a behavioral signal, not a strict token budget, on every model. If you want a shorter answer, ask for one.
3. Changing effort mid-conversation the old way. On most models, changing effort between requests starts your prompt cache over, and in a long session that is expensive. On Fable 5.1, Opus 5.5, Opus 5 and Sonnet 5.5, the API now offers a per-message effort change, in beta, that keeps the cache (not with Sonnet 5.5's between_tools setting). In Claude Code, changing effort keeps the cache on Opus 5.5, Sonnet 5.5 and Fable 5.1 when you sign in with an API key or a Claude subscription. Anywhere else, pick a level at the start and hold it.
4. Running xhigh or max with a small output limit. At those levels Claude needs room to think, and thinking counts against your output limit alongside the response. For Opus 5, 4.8 and 4.7, Anthropic suggests starting at 64k tokens and tuning from there. For agentic coding on Sonnet 5.5, it says to set 128,000, the model's maximum.
5. Trying to switch thinking off. In the API, sending thinking disabled returns an error at every effort level on Fable 5.1, Opus 5.5 and Sonnet 5.5. Sonnet 5.5 offers between_tools instead, at high or below. On Opus 5 it returns an error at xhigh and max. In the Claude app, thinking cannot be turned off on Sonnet 5.5, Opus 5.5, Fable 5.1 or Opus 5.
Where you set this
In the Claude API, you set effort on each request. The default is high on every model that has effort, except Opus 5.5, which defaults to medium.
In Claude Code, the default model on Pro, Max, Team, Enterprise and the API is Opus 5.5. The default effort is medium on Opus 5.5 and Sonnet 5.5, and high on most other models. Change it with /effort, or with the slider in /model.
In the Claude app, the effort control sits next to the send button: click the model name, then Effort. Each model's recommended level is marked Default in that menu. Anthropic's help article does not list which level that is for each model, so check in the app rather than assuming. The one app default Anthropic has published is Sonnet 5.5 at medium.
If you use Claude only through the chat app, both decisions on this page are now yours to make there too.
Verification
Every figure and recommendation on this page was checked on September 30, 2026 against:
platform.claude.com/docs/en/models/overview for models, pricing, context, cutoffs, and default effort
platform.claude.com/docs/en/about-claude/pricing for prices
platform.claude.com/docs/en/about-claude/models/choosing-a-model for Anthropic's starting advice
platform.claude.com/docs/en/build-with-claude/effort for effort levels, per-model support, and guidance
platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting for which models reject thinking
disabledplatform.claude.com/docs/en/about-claude/model-deprecations for legacy status, retirement dates, and notice periods
code.claude.com/docs/en/model-config and code.claude.com/docs/en/prompt-caching for Claude Code defaults and cache behavior
support.claude.com, "Change the model, effort, and thinking settings," for the Claude app
anthropic.com/claude-opus-5-5 and anthropic.com/claude-sonnet-5-5, Anthropic's launch pages
Model lineups have been moving every few weeks. Opus 5.5 shipped on September 22 and Sonnet 5.5 on September 28. Effort guidance moves with each model release.
Next verification: November 28, 2026.
Eric Melillo writes about AI as operator leverage at ericmelillo.com/newsletter. Twenty years scaling other people's companies, two agencies built and exited, now rebuilding the same systems for operators working alone.