Claude Sonnet 5.5 in Claude Code: The Alias Moved on One Provider, and the Default Moved Down

The third-party briefing that put Claude Sonnet 5.5 back on the radar this week said it is "live everywhere, including the Platform and Claude Code." The first half is the vendor's own line from September 28, and it checks out: Claude Sonnet 5.5 shipped on the Claude API, Amazon Bedrock, Claude Platform on AWS, Goog​le Cloud and Microsoft Foundry at $2 and $10 per million tokens, unchanged from Claude Sonnet 5. The second half is true only if "Claude Code" means one provider. The Claude Code build that landed the same afternoon, v2.1.284, points the sonnet alias at Claude Sonnet 5.5 on the Anthro​pic API — and leaves it pointing at Claude Sonnet 4.6 on Claude Platform on AWS and Claude Sonnet 4.5 on Amazon Bedrock and Goog​le Cloud. Run /model sonnet against a Bedrock deployment and you do not get the new model. There is a second thing that moved, and almost nobody has written it down: on the Claude API Claude Sonnet 5.5 starts at high effort; in Claude Code it starts at medium. Both facts are documented, both are load-bearing for anyone deciding whether to switch this week, and both sit underneath the same sentence about the model being live everywhere.

What v2.1.284 actually changed

The Claude Code changelog entry is one line long and the important half is a qualifier: "Added Claude Sonnet 5.5 (claude-sonnet-5-5), now the default Sonnet model on the Ant​hropic API — 1M context, $2/$10 per Mtok with $0.20/Mtok cache reads." That build published to npm on September 28 at 17:11 UTC, which is the release date with a timestamp on it rather than a version number. The model-configuration documentation carries the resolution table the changelog does not.

• sonnet alias — Claude Sonnet 5.5 on the Ant​hropic API; Claude Sonnet 4.6 on Claude Platform on AWS; Claude Sonnet 4.5 on Amazon Bedrock and Goo​gle Cloud's Agent Platform; Claude Sonnet 4.5 on Microsoft Foundry.

• opus alias — Claude Opus 5.5 on the Ant​hropic API, Claude Platform on AWS, Amazon Bedrock and Goo​gle Cloud; Claude Opus 4.6 on Microsoft Foundry.

• default model — Claude Opus 5.5 on Pro, Max, Team, Enterprise and the Ant​hropic API, and Claude Sonnet 4.5 on Microsoft Foundry. It has been Opus 5.5 since v2.1.280 on September 22; before that it was Claude Sonnet 5 on Pro and Team Standard.

• sonnet[1m] — no effect where sonnet already resolves to Claude Sonnet 5.5 or Claude Sonnet 5, because both carry a native 1M-token window. The suffix is not how you get long context on this model; the model already has it.

Read the first and third rows together and the practical consequence is narrow but sharp. On the Ant​hropic API, upgrading Claude Code is the whole migration — sonnet moves for you. On Bedrock, Goo​gle Cloud and Claude Platform on AWS, the model is served but the alias is not moved, and the documentation is explicit about the workaround: where an alias resolves to an older model, newer models are reachable by naming them in full or by setting ANTHROPIC_DEFAULT_SONNET_MODEL. So "Sonnet 5.5 is live in Claude Code" is a sentence about a default, and defaults are per provider.

The default effort is a different rung on each surface

This is the part that changes your bill rather than your model string, and it is split across two pages that do not cross-reference each other. The Claude API's effort documentation states that Claude Sonnet 5.5 supports all five levels and that high is the default there. The Claude Code model-configuration page states the resolution order for a session — an explicit choice (the CLAUDE_CODE_EFFORT_LEVEL variable, --effort, or /effort), then your saved settings, then the model's own default — and then gives that model default as: "high on every model that supports effort, except that Opus 5.5 and Sonnet 5.5 default to medium, Opus 4.7 defaults to xhigh."

Two defaults, one model, one rung apart. On the independent index that one rung is worth about six intelligence points and roughly 84% more money per completed task. Ant​hropic's own guidance for Claude Sonnet 5.5 — start at high unless the workload is agentic or latency-sensitive, and start at medium for agentic coding and multistep tool use — is written for the API surface, where high is what you get by doing nothing. Claude Code has already made that call for you in the other direction.

Whether it made the right call is genuinely workload-dependent, and the vendor's argument for medium is defensible: the model's effort levels are explicitly recalibrated against Claude Sonnet 5, and the migration note says to re-run an effort sweep rather than carry a setting over. But "recalibrated" cuts both ways, and the next section is where it lands.

What the ladder costs, measured

Artificial Analysis has Claude Sonnet 5.5 measured at every rung on Intelligence Index v4.3.2, which is a ten-evaluation suite, and the per-rung cost is the number a rate card cannot show you. All figures below are that evaluator's, read from the model page's own data on September 30, 2026; the vendor rates are Ant​hropic's.

• Low — index 35.84 at $0.41 per task; first chunk in 1.04 seconds.

• Medium, the Claude Code default — 40.74 at $0.59; first chunk in 1.34 seconds.

• High, the Claude API default — 46.74 at $1.08; first chunk in 16.6 seconds.

• Xhigh — 51.85 at $2.74; first chunk in 39.7 seconds.

• Max — 55.98 at $7.60; first chunk in 378.8 seconds.

Now put Claude Sonnet 5's own ladder beside it, because that is the model most readers would be migrating from. Its rungs read 24.26, 28.05, 31.66, 34.38 and 38.16, costing $0.51, $1.00, $1.79, $2.87 and $5.09 per task. Claude Sonnet 5.5 at the Claude Code default of medium scores 40.74 — above Claude Sonnet 5's best result at any effort level — for $0.59 against $5.09. That is 8.7 times cheaper per completed task on the same suite, and it is the independent version of Ant​hropic's "less than a tenth of the cost" claim, which the vendor makes on Terminal-Bench at Medium effort. Same direction, slightly smaller factor, and it is a measured result rather than a vendor table. Drop to low instead and the ordering flips: 35.84 is below Claude Sonnet 5's ceiling, which is the honest floor on the saving.

The rung that is not cheap on either model is the top one. At max effort Claude Sonnet 5.5 costs $7.60 per index task against Claude Sonnet 5's $5.09 — about 49% more, not less. That is not a contradiction of the vendor's number. It is the same curve read at the other end, and it is where a migration lands if the effort setting is carried over rather than swept.

Your saved effort setting may not follow you

There is a third setting in play and it is the one most likely to be silently wrong. Claude Code saves effort per model, under modelSettings in your user settings, so each model keeps its own level. The older form — a single top-level effortLevel key, which is what /effort wrote before per-model saving existed — is handled differently for recent models. The documentation says that key "keeps applying where it applied before, on Opus 5, Fable 5.1, and earlier models, while Opus 5.5 and models released after it start at their own default until you choose a level for them."

Claude Sonnet 5.5 was released after Claude Opus 5.5 — September 28 against September 22 — so it falls in that second group. A team that set a global effortLevel: "high" a year ago to keep Claude Sonnet 5 sharp does not get high on the new model. They get medium, which is the Claude Code default, and nothing in the session says so. The same page notes that max applies to the current session only unless it is set through the environment variable, and that project, local or managed effortLevel keys, or one passed with --settings, do apply to every model. If you want an effort policy that survives a model release, that is the form to write it in.

One adjacent change is worth naming because it shares the version number: the Ultracode toggle in the effort slider is a Claude Code setting rather than a model effort level, and keeping it on at levels other than xhigh, or turning it off with /effort ultracode off, requires v2.1.284 — the same build that carried the new Sonnet. Ultracode is not an effort level and should not be read as one, but if you scripted around the old behaviour, that is the version where it changed.

Waiting is the thinking budget, not the token rate

The latency figures above are the most operationally useful thing on the page, because they invert the usual intuition. Output speed on Claude Sonnet 5.5 does not fall as effort rises; it rises with it, from 85 tokens per second at low effort to 137 at max. What explodes is the wait before the first token exists at all — 1.0 second at low, 1.3 at medium, 16.6 at high, 39.7 at xhigh, 378.8 at max. The model is not slower per token at the top of the ladder. It is thinking for longer before it starts, and at the top rung that is more than six minutes of silence in front of every task.

That reframes what the setting is for. medium is not a quality compromise so much as the only rung where the model behaves like a conversational tool; max is a batch operation wearing an interactive interface. A streaming client that hides the gap makes high look fine in a demo and feel broken in a terminal. The practical split most teams will land on is a low rung for anything a human is watching and a high one for work that runs unattended — which is close to the vendor's own advice, arrived at from the latency side rather than the cost side.

The comparison that is actually routable today

If the point of the ladder is that medium is where the value sits, the interesting question is what else sits near it. Claude Opus 5.5 at its own Claude Code default of medium scores 51.24 at $1.34 per task on the same suite. Claude Sonnet 5.5 at high scores 46.74 at $1.08. Four and a half index points separate them for about a quarter more per task, and Claude Opus 5.5 is measurably ahead at every rung above its default as well.

One availability note, stated rather than implied: Claude Sonnet 5.5 is not in the OrcaRouter catalogue. It is absent from the public model list and every spelling of the slug returns a not-found, so nothing here should be read as a route to it. What is routable is the model you would be migrating from and the model that sits above it. anthropic/claude-sonnet-5 has been on the platform since June 30, 2026 at Ant​hropic's own $2.00 and $10.00 per million tokens with $0.20 cache reads and a $2.50 cache write, served over /v1/messages, /v1/chat/completions and /v1/responses. anthropic/claude-opus-5.5 has been there since September 22 at $4.00 and $20.00, over the Ant​hropic-compatible and chat-completions endpoints.

That combination is what makes an effort sweep cheap to run. Both sides of the comparison — the incumbent at its old setting and the larger sibling at the rung the new model's default is being weighed against — answer on one credential with no second contract and no code change beyond the model string, and the catalogue passes provider list price through without adding a markup. The one thing it will not do is serve you Claude Sonnet 5.5 itself, and until that changes, the honest version of this article's advice is to evaluate the rung on Ant​hropic's own API and evaluate the alternatives where they are already reachable.

What to do this week

Three things, in the order they will cost you money if you skip them. Check which provider your Claude Code session actually runs against before you assume the upgrade landed, because on Bedrock, Goo​gle Cloud and Claude Platform on AWS the sonnet alias still points at Claude Sonnet 4.5 or Claude Sonnet 4.6 and the model has to be named in full. Check whether your saved effort is a per-model modelSettings entry or the older top-level key, because only the first one follows you onto a model released after Claude Opus 5.5. And when you re-run the sweep the vendor asks for, sweep upwards as well as down: at the Claude Code default Claude Sonnet 5.5 beats Claude Sonnet 5's best result for under a ninth of the cost, which means the interesting question is no longer whether to move down, but how far up you can afford to go.

The open question is Claude Haiku 5.5, which Ant​hropic says will join the family in the coming weeks and will sit underneath this model. If the pattern from Claude Opus 5.5 holds, it will arrive on the Ant​hropic API first and work its way outward to the other providers on a delay measured in Claude Code releases rather than days. That delay is the thing to watch, because everything above about aliases and defaults is really a consequence of it.