Sonnet 5.5 is near-Opus at half the price

Anthropic shipped Claude Sonnet 5.5 on Monday, and the headline is not the price. The tariff stayed exactly where Sonnet 5 left it: $2 per million input tokens, $10 per million output, $0.20 per million cache reads. The headline is what that money now buys.

On Anthropic's own Terminal-Bench 4.0 evaluation, Sonnet 5.5 scores 70.6%, ahead of Opus 5.5's 66.4% and up from Sonnet 5's 10.3%. On the company's GDPval-AA knowledge-work benchmark it lands two points off the flagship: 1844 against Opus 5.5's 1846. CursorBench is closer still, 55.5% versus 57.8%. Every one of those numbers is self-reported, and the fine print matters (more on that below). But even with a heavy discount applied for vendor optimism, this is a mid-tier model posting flagship-adjacent numbers at half the flagship's per-token price. Opus 5.5 bills $4/$20. Do the division yourself.

This lands hardest on teams that just re-routed work to Opus 5.5 after last week's price cut. If you moved agentic coding loops to Opus on cost grounds, the case has shifted again, only a week later. The default for well-scoped work, bug fixes, and coding-agent steps should now be Sonnet 5.5 until your own numbers say otherwise. Anthropic's own steering agrees: it calls Sonnet 5.5 the model for "well-scoped everyday tasks" and reserves Opus for work needing sustained judgment.

Here is where the skepticism earns its keep. Anthropic claims Sonnet 5.5 costs "up to 30% less per task" than its predecessor in the company's testing. The one independent dataset available disagrees at the top end. Artificial Analysis, via the Van Data Team's benchmark teardown, measured Sonnet 5.5 at max effort costing $7.60 per task on its Intelligence Index: more than Sonnet 5's $5.09 at the same setting and more than Opus 5.5's $5.98, because Sonnet 5.5 at max "used the most output tokens per task it had seen." At low, medium, and high effort, Sonnet 5.5 beats Sonnet 5 on both score and cost, exactly as Anthropic claims. At max effort the story flips. Opus 5.5 at high effort even outscores Sonnet 5.5 at xhigh (54 to 52) for about a third less money ($1.82 against $2.74).

So the routing rule is not "Sonnet is cheaper now." It is: cheaper per token, cheaper per task at low through high effort, and actively worse value at max. Measure cost per completed task on your workload before you touch the max slider. This is the same "measure the task, not the token" lesson from the price-war comparison, now with a real reversal to point at.

GitHub, meanwhile, did not wait. Sonnet 5.5 is generally available in Copilot as of the same day, per GitHub's changelog, billed at provider list pricing under usage-based billing across Pro, Pro+, Max, Business, and Enterprise. The rollout is gradual, and under default model enablement the new model switches on automatically unless an admin opted out. Copilot admins should check their model policy if they gate new models.

The bigger story for agent operators is not pricing at all. Sonnet 5.5 is the first Sonnet to ship with Fable-style safeguards, and that changes what you should expect in production.

First, cyber safeguards. Because Anthropic rates Sonnet 5.5's cybersecurity capabilities as comparable to Opus 5's, it launches with safeguards like those on its most capable models. Routine bug finding and fixing are unaffected, the company says. Higher-risk cybersecurity tasks "will visibly fall back to Sonnet 5." Second, anti-distillation. Sonnet 5.5 is the first Sonnet with safety classifiers that block reasoning extraction, and it expands preserved thinking: its thinking blocks are bound to the account that produced them, and a block replayed from another account gets dropped before the model sees it. The request still succeeds, just without the thinking context.

Both of these have sharp edges for harness builders. Anthropic has not published what counts as a "higher-risk cybersecurity task," so security teams cannot know in advance which prompts will fall back. And the account-bound thinking means moving a conversation between accounts, including switching accounts mid-session in Claude Code, now silently loses reasoning context. The release notes are explicit: the API drops the block and succeeds. If your pipeline assumes a model fallback or a context move is invisible, instrument it. Log the serving model on every response. That fallback changes quality and cost mid-task, and nothing in the docs tells you whether the response payload says which model actually answered.

The migration checklist is short but real. On 5.5, thinking: {"type": "disabled"} is gone; send between_tools instead (up to high effort). Forced tool use (tool_choice set to any or tool) returns a 400. The old computer_20251124 tool is rejected on the Claude API and Google Cloud. Effort defaults to Medium in Claude Code and the apps, High on the platform. None of this is a surprise to anyone who watched the Fable-class rollouts, but it is new to the Sonnet tier, and teams that treat a Sonnet-to-Sonnet upgrade as a drop-in swap will find out the hard way.

What remains unknown: the fallback criteria, whether the Cyber Verification Program (promised as expanded "soon") opens on a timeline that matters, whether the 30%-less-per-task claim holds outside Anthropic's harness, and what Haiku 5.5, due in the coming weeks, does to the bottom of the tariff ladder. Also worth remembering: the refusal-billing change from September 24 applies to Sonnet 5.5 too. Refusals in the bio, frontier_llm, and reasoning_extraction categories now bill like any other request, at this model's rates. The cyber safeguards that are new to Sonnet are the same classifiers that can refuse, so budget for the edge case.

One platform footnote while the release notes are open: cache diagnostics is now generally available on the API without a beta header, and the Compliance API's local session endpoints are out of beta for Claude for Microsoft 365 Office surfaces. Unrelated to Sonnet 5.5, but landing in the same notes, and both are the kind of small capability you only notice when you need it.

The pattern of the month keeps repeating. Anthropic ships a cheaper or faster model, the vendor chart looks fantastic, the tariff arithmetic checks out, and then the per-task reality is more complicated than the launch post. Sonnet 5.5 is a strong model at a fair price. Run your own evals, route by task and effort level, instrument the fallbacks, and read the safeguards section twice.