Anthropic's Own Docs Say You Are Probably Picking the Wrong Claude Model
Four Claude models, a 10x price spread, and a default recommendation that is not the most expensive one. The selection guide also says tuning effort beats switching models.
The WJS Desk
Sep 1, 2026 · updated 4 hours ago · 6 min read
There are four Claude models you can call today, and the gap between the cheapest and the most expensive is ten times on output tokens: $5 per million for Haiku 4.5, $50 for Fable 5. That spread is large enough that picking wrong is a budget decision rather than a preference.
So we read Anthropic's own documentation to see what it actually tells you to do. Two things stood out. Its default recommendation is not the most expensive model. And buried in the selection guide is a line that reframes the whole question:
Tuning effort is often a better lever than switching models.
We did not run our own benchmarks for this piece. Every number below comes from Anthropic's published documentation, and every opinion from a named developer is linked to where they said it.
The lineup, from the source
These are the four current models and their published specifications. Prices are per million tokens:
| Haiku 4.5 | Sonnet 5 | Opus 5 | Fable 5 | |
|---|---|---|---|---|
| Input / output | $1 / $5 | $2 / $10 | $5 / $25 | $10 / $50 |
| Context window | 200K | 1M | 1M | 1M |
| Max output | 64K | 128K | 128K | 128K |
| Relative latency | Fastest | Fast | Moderate | Slower |
| Knowledge cutoff | Feb 2025 | Jan 2026 | May 2026 | Jan 2026 |
| Effort parameter | Not supported | Yes | Yes | Yes |
| Retires no sooner than | 15 Oct 2026 | 30 Jun 2027 | 24 Jul 2027 | 9 Jun 2027 |
Two adjustments apply to all of them. Batch API requests are 50% off, and prompt cache reads cost 10% of the base input price. If your workload is repetitive or tolerant of latency, those two levers move the bill further than a tier change does.
Anthropic's stated default is unambiguous, and it is the middle of the range rather than the top: start with Opus 5 for complex agentic coding and enterprise work, and reach for Fable 5 only when you need "the highest available capability."
The advice nobody quotes
The selection guide offers two starting strategies. Efficiency-first means beginning on Haiku 4.5 and upgrading only where you hit a capability gap. Capability-first means starting on Opus 5, optimising prompts, then stepping down. Both are reasonable, and which suits you is mostly a question of whether accuracy or unit cost is the thing that will kill the project.
But the more useful line is the one about effort. Recent Opus and Sonnet models expose an effort parameter that trades intelligence against latency and cost within a single model. Anthropic's guidance is to treat that as the first lever, not the last: on Opus 5, start at the default high and adjust in either direction based on your evals.
That matters because "which model?" is a question people ask once and answer permanently, while effort is a dial you can turn per workload. A team that moved from Opus 5 to Sonnet 5 to halve costs might have got most of the saving by lowering effort and keeping the model.
More effort is not reliably better. Reading the Opus 5 system card, Hacker News commenter Dibes flagged a graph showing medium as the most effective thinking mode by far for frontier code, calling it "the only case that I saw going through the system card where more reasoning effort meaningfully negatively impacted the resulting eval." If you assume the top setting is the safe default, that assumption is worth testing on your own tasks.
Two things that decide it and appear on no benchmark chart
Data retention. Prompts sent to and outputs from "covered models" are retained for 30 days to support Anthropic's safety work, on every platform where those models are offered. Fable 5 and Mythos 5 are covered. Opus 5, Sonnet 5 and Haiku 4.5 are not. Commenter postalcoder argued this is the real story of the Opus 5 release, not the benchmarks: organisations now get near-Fable capability without Fable's retention requirement. If you have a zero-retention agreement or a compliance team, that single row can decide the tier for you regardless of what any leaderboard says.
Retirement dates. The cheapest model is also the one closest to the exit. Haiku 4.5 retires no sooner than 15 October 2026, roughly six weeks from now. Every other current model has a date in 2027. Anthropic commits to at least 60 days' notice before retiring a publicly released model, and the dates are "not sooner than" rather than fixed, so this is not a cliff. But if you are choosing Haiku today for a system meant to run for a year, you are choosing a migration.
Haiku carries a second cost that is easy to miss: its reliable knowledge cutoff is February 2025, against May 2026 for Opus 5. For a summariser working from supplied context that is irrelevant. For anything reasoning about current libraries or APIs, it is fifteen months of drift.
Worth checking if you are on a cloud platform: these dates apply to Anthropic-operated platforms. Amazon Bedrock and Google Cloud set their own retirement schedules, so the model you are calling may have a different lifecycle than the documentation above implies.
The gotcha that returns a 400
If you are migrating older code, this one bites immediately. temperature, top_p and top_k are deprecated from Claude Opus 4.7 onward, and setting any of them to a non-default value returns a 400 error rather than being quietly ignored. Anthropic's recommendation is to drop them and steer behaviour through prompting instead.
The Python SDK goes further: from v1.0 it removes those parameters from its request types entirely, so passing them raises a TypeError before a request is ever made. Most other SDKs keep them in the types so existing code still compiles, which means the failure surfaces at runtime instead.
What developers are actually arguing about
The Opus 5 launch thread drew 1,778 points and 1,335 comments, and the recurring complaint was not capability. It was that the choice has become unmanageable. paxys put it best:
There are 10+ LLM companies, each with dozens of models of different modalities, each model with multiple size variants, then different "thinking" levels, then agentic modes, "pro" modes, a "fast" option, standard vs flex vs batch execution.
Their conclusion, that model routing is the fastest growing segment in AI right now, follows fairly directly from a four-model lineup where the second-cheapest frontier option is the recommended default.
The positioning confuses people too. 6thbit noted that Anthropic says Opus 5 is not more capable overall than Fable 5, while the launch post lists Opus 5 ahead on most benchmarks shown. theHocineSaad asked the obvious follow-up: if Opus 5 scores higher at half the price, why is Fable 5 still the flagship?
We think the answer is in the docs rather than the marketing. Fable 5 is described as next-generation intelligence for long-running agents, with adaptive thinking always on. Benchmarks are mostly short tasks. A model tuned for multi-hour autonomy is not the model that wins a one-shot eval, and Anthropic's own selection matrix reserves Fable for long-horizon agentic work rather than for scoring well.
How we would actually decide
Start where Anthropic says to start, which is Opus 5 for real work and Haiku 4.5 if you are prototyping or latency-bound. Then, before you consider changing tier, turn the effort dial and measure. Then batch what can be batched and cache what repeats, because 50% and 90% discounts respectively beat any tier change you were contemplating.
Check the two non-benchmark rows before you commit: if your data cannot sit anywhere for 30 days, Fable is out regardless of capability, and if you are building something meant to outlive the year, Haiku's October date is a scheduled migration rather than a footnote.
The most expensive model is not the default, the top effort setting is not always the best one, and the cheapest tier expires first. All three are in the documentation and none are in the marketing.
The one piece of advice worth repeating is Anthropic's own, and it is the least exciting sentence in the whole selection guide: having a good evaluation set is the most important step in the process. Every comparison in this article, ours included, is a proxy for the only test that matters, which is your prompts on your data.
What would change our read: Anthropic publishing standardised latency figures, which it currently does not, or Haiku 4.5 getting a successor before October. The second is the one to watch, because a Haiku 5 would reset the cheapest tier's clock and remove the only real argument against starting there.
