Claude Sonnet 5.5 Beats Opus at Coding, Half the Price

Photo: Developers coding at the Junction 2015 hackathon. Photo: Vmuru, via Wikimedia Commons (CC BY-SA 4.0). Photo from 2015; it does not depict Anthropic's model.

Six days after shipping its flagship Opus 5.5, Anthropic has done something it has never done before: released a mid-tier model that beats its own most expensive one at coding — for half the price.

Anthropic released Claude Sonnet 5.5 on September 28, the second model in what it calls the Claude 5.5 family after Opus 5.5's September 22 debut. The announcement came via the @claudeai account on X: "Introducing Claude Sonnet 5.5... It's a clear upgrade over Sonnet 5, runs more than 30% faster, and costs up to 30% less for most work."

The headline numbers are striking. Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0, a test of whether an AI agent can complete complex professional tasks from the command line — against 66.4% for the pricier Opus 5.5 and just 10.3% for its predecessor Sonnet 5. On GDPval-AA, a benchmark built on real professional work, it lands at 1844 Elo — two points behind Opus 5.5's 1846, and light-years ahead of Sonnet 5's 1449. Independent testing firm Artificial Analysis confirms the story, ranking Sonnet 5.5 ahead of Opus 5.5 on Terminal-Bench, 63.6% to 59.6%, with OpenAI's GPT-6 Astra trailing at 59.1%.

And yet the list price is unchanged from Sonnet 5: $2 per million input tokens and $10 per million output tokens — exactly half of Opus 5.5's $4/$20. Cache reads cost $0.20 per million tokens, cache writes $2.50. The model is available as claude-sonnet-5-5 across the Claude Platform, AWS, Google Cloud, and Microsoft Azure, with zero data retention. A smaller Claude Haiku 5.5 is promised in the coming weeks.

The benchmarks that matter

Vendor numbers always deserve scrutiny, so here is the full table as published in Anthropic's system card. Terminal-Bench 4.0: 70.6% (Opus 5.5: 66.4%, Sonnet 5: 10.3%). CursorBench 4.0: 55.5%, about two points below Opus 5.5's 57.8%. FrontierCode 1.1: 52.1% at Xhigh effort and 46.2% at Max — against 49.3% for OpenAI's GPT-6 Sol. OSWorld 2.1, a computer-use test: 80.1%, close to Opus 5.5's 81.8%. Humanity's Last Exam with tools: 64.5%, up from 54.9% on Sonnet 5.

The most important number is not on the list price at all. Anthropic says Sonnet 5.5 needs nearly a third fewer tokens per task than its predecessor, which is why it costs up to 30% less per task at the same list price. Customer data backs the claim: Balyasny Asset Management measured about 121,000 tokens per answer versus 497,000 on Sonnet 5; Base44 reported 3.6 iterations per app build versus 7.7 on Opus 5; Zendesk says it processed tickets 20% faster. Slack principal engineer Curtis Allen said Sonnet 5.5 beat Sonnet 5 on nearly all of Slack's offline Slackbot evaluations, finishing in fewer steps with about 14% fewer output tokens.

"Sonnet 5.5 is strongest at well-scoped everyday tasks, fixing bugs, and creating polished documents, slides, and spreadsheets. It's also got a sharp eye for design," Anthropic wrote in its announcement. And crucially: "Where Opus 5.5 is built for complex work requiring careful judgment," the company says, Opus remains the model for hard, open-ended work. Anthropic is not pretending the gap has closed entirely — it is telling buyers exactly where to draw the line.

The catch hiding in the effort dial

But the pricing story has a sharp edge, and independent testers found it. Sonnet 5.5 ships with an effort setting — low, medium, high, xhigh, and max — that controls how long the model thinks before answering. The headline benchmarks run at high-to-max effort, and the cost curve bends hard at the top.

Artificial Analysis measured Sonnet 5.5 at roughly $7.60 per task at maximum effort — above Opus 5.5's $5.98, and a reversal of the price advantage the list price implies. The model used more tokens per task than any model the firm tested. As one analyst put it, at the same $2/$10 rate as its predecessor, the effort setting matters more than the price list.

Anthropic is not hiding this: it published cost-per-task curves for every effort level. And the company notes that at High effort, Sonnet 5.5 matches GPT-6 Sol on FrontierCode for about a fifth of the cost per task. The default effort differs by surface — Claude Code and the Claude apps default to Medium, the Claude Platform to High — so most developers will never touch the dial. But teams running maxed-out agent fleets should budget with the effort curve, not the price card.

The first Sonnet with frontier-grade safeguards

Here is the part of the release that got less attention than the benchmarks but may matter more. Sonnet 5.5's cybersecurity capabilities are, in Anthropic's words, a large improvement over Sonnet 5's and comparable to Opus 5's — which is why it is the first Sonnet model to ship with cyber safeguards similar to those on Opus 5.5. Routine software development and bug fixing are unaffected; higher-risk cybersecurity requests will visibly fall back to Sonnet 5, with products telling users when that happened. Biology safeguards are unchanged from Sonnet 5, and an expanded Cyber Verification Program will soon offer vetted security teams tiered access.

Sonnet 5.5 is also the first Sonnet to launch with anti-distillation classifiers designed to block reasoning extraction — the technique by which bad actors use thousands of fake accounts to extract a model's capabilities at industrial scale and train rival models without its safeguards. Anthropic also expanded "preserved thinking," so Claude's reasoning cannot be decoupled from the account that created it. The practical consequence: conversations can no longer move between accounts, and Claude Code sessions break if you switch accounts mid-session.

On alignment, Anthropic ran an automated behavioral audit of roughly 1,850 scenarios: Sonnet 5.5 matched or improved on Sonnet 5 on most alignment measures, with Opus 5.5 still performing slightly better overall. The company says Sonnet 5.5 does not extend its capability frontier, so its alignment review focused on risks such as misleading users and aiding high-stakes misuse — and it acknowledges plainly that no set of evaluations catches every failure.

The model wars enter a new phase

Step back and the week tells a story about two labs moving in opposite directions. Anthropic shipped Opus 5.5 on September 22 and Sonnet 5.5 six days later, accelerating its release cadence. On September 28 — the same day — OpenAI scrapped the release of its newest model, GPT-6.1 Astra, after it fell short of the company's safety standards, a decision CNBC confirmed on Monday.

That contrast is the real news here. One lab is slowing down because its models scare it; the other is speeding up and wrapping the result in the first frontier-grade safeguards a mid-tier model has ever carried. Anthropic is betting that the market wants capability with guardrails baked in — cyber fallbacks, anti-extraction classifiers, preserved thinking — rather than capability held back entirely.

The economics favor the bet. If Sonnet 5.5 genuinely delivers near-Opus coding performance at half the list price and meaningfully lower per-task cost at sensible effort settings, it becomes the default model behind a generation of developer tools — and the pricing signal ripples straight into OpenAI's and Google's product planning. Anthropic's second 5.5 model in a week, with a third — Haiku 5.5 — on the way, is a deliberate attempt to own the efficiency tier before anyone else defines it.

Who should switch

The honest buying guide, per Anthropic's own data: if your workloads are well-scoped everyday coding, bug fixing, and document work, Sonnet 5.5 is now the model to start with — at High effort or below. If your work needs sustained judgment on complex, open-ended problems, Opus 5.5 still wins, and Anthropic says so itself. If you run agents at max effort, watch the token burn before you rewrite your routing logic; independent verification on FrontierCode and CursorBench matters more than any vendor table.

But zoom out from the dials and the price card, and this release marks something larger: the moment the mid-tier stopped being the cheap compromise and became the frontier. Anthropic's cheapest serious model now out-codes its flagship. The price-performance curve of AI just snapped — and the snap happened six days after the company's flagship shipped.

Related reading: the other side of the race — "OpenAI Scraps GPT-6.1 Astra Over Safety Fears on DevDay Eve."

Sources: Anthropic (Sept. 28, 2026 — Sonnet 5.5 announcement and system card, primary); @claudeai, X (Sept. 28, 2026); MarkTechPost (Sept. 28, 2026 — benchmark table, customer data); Artificial Analysis (Sept. 28, 2026 — independent Terminal-Bench testing, cost-per-task figures); Gadgets Now (Sept. 29, 2026); Mashable Light Speed (Slack evaluation figures); Fone Arena (Sept. 28, 2026 — safeguards); The Deep View (Sept. 28, 2026). Benchmarks are vendor-reported unless attributed to Artificial Analysis; per-task cost figures at max effort are from independent testing.