Claude Opus, Sonnet and Haiku Are All 5.5. Fable is still 5.1.
Anthropic moved three model tiers to 5.5 in sixteen days. The flagship did not move, and its own docs now tell you to start with the cheaper model.
Between September 22 and October 7, Anthropic moved three of its four model tiers to 5.5. Opus on the 22nd, Sonnet six days later, Haiku nine days after that. Sixteen days, three releases.
The fourth tier did not move. Claude Fable, the flagship, is still 5.1, shipped on September 1 and untouched since.
16 days
From Opus 5.5 to Haiku 5.5. Three tiers, one version number.
5.1
Where Fable has sat since September 1, while the other three moved past it.
2.5x
What Fable 5.1 costs per token against Opus 5.5. The flagship is the expensive one.
The instruction buried in the docs
Anthropic's own model overview opens with routing advice, and it is not the advice you would expect a lab to give about its flagship. It says to start with Claude Opus 5.5 for most workloads, and to use Claude Fable 5.1 for demanding reasoning and long-horizon agentic work, "or when your evals on Claude Opus 5.5 at higher effort still fall short."
Read the second half of that sentence again. The flagship is documented as the thing you reach for after the default has failed your evals. That is not how a halo product is described. That is how a specialist is described.
The cheaper model already won a benchmark
This is not just positioning. On Anthropic's own launch pages, Sonnet 5.5 scored 70.6% on Terminal-Bench 4.0 and Opus 5.5 scored 66.4%. The model that costs half as much beat the one documented as the default, on the agentic coding evaluation, in the lab's own published numbers.
Anthropic did not hide it. It is in the table. But it tells you something about what "flagship" now means: not the model that wins, the model that costs the most to run and is held for the cases nothing else clears.
You can see the same thing in how people actually work. Karpathy built HackerTube in about two hours and named the model he used. It was Opus 5.5, not the flagship.
The four tiers are now the same machine at four prices
Look at what no longer separates them. All four carry a 1M token context window, all four cap output at 128K, and all four share a June 2026 knowledge cutoff. The capabilities table that used to justify the ladder has flattened: the differences left are price, latency, and how much thinking each does by default.
Which turns tier selection into a budgeting exercise rather than a capability one. You are not choosing what the model can see or how recently it learned. You are choosing how much you are willing to pay per token and how long you will wait, and those are questions your finance team can answer as well as your engineers can.
It also turns the choice into a routing problem rather than a standards decision, which is why that layer keeps growing. OpenRouter measured 194 trillion tokens in a single week precisely because nobody wants to hardcode one tier, let alone one lab, while the prices and the leaderboard keep moving.
Why people keep announcing Fable 5.5
On November 2 two accounts announced Claude Fable 5.5, one of them a harness shipping support for it. There is no such model. Anthropic's newsroom has no announcement, the docs list no such version, and the API has no such ID. The newest Fable is 5.1.
That is not really a mistake about Anthropic. It is a prediction being reported as a fact, and the prediction is reasonable: three tiers moved, so the fourth looks overdue. The naming makes the gap feel like an error waiting to be corrected.
It might be. Or the number stays where it is because the tier is no longer on the same ladder. A model you reach for when your evals fail does not need to track the version number of the model whose evals failed.
What to do with four models and three prices
Default to Opus 5.5 and make Fable earn its place. Anthropic is telling you this in its own documentation and the price difference makes it expensive to ignore. Fable 5.1 is 2.5 times the cost per token, and that is before you account for it running at high effort by default, which it does.
Run the eval before you pay the premium. The docs give you the exact trigger: Opus 5.5 at higher effort, failing your tests. If you have not run that, you do not yet know whether you need the flagship. Most teams reaching for the top of the menu have never checked the rung below it.
Check Sonnet on agentic work specifically. The Terminal-Bench result is one benchmark and should not be over-read, but it points at a real question. If your workload looks like long tool-using sessions rather than hard reasoning, the mid tier may be the right default and it costs a quarter of the flagship.
Pin your version, and watch the retirement dates. All four IDs are pinned snapshots now, with no dateless alias resolving underneath you. Fable 5.1 retires no sooner than 1 September 2027 and Haiku 5.5 no sooner than 7 October 2027. That is your planning horizon, published.
Three tiers moved in sixteen days. The one that stayed still is the one most people assume they should be using.
Start one rung lower.


Fable 5.1