Anthropic shipped Claude Sonnet 5.5 on September 28 and left the price alone: $2 per million input tokens, $10 per million output, same as Sonnet 5. The score moved instead. On GDPval-AA v2.1, one of the benchmarks in Anthropic’s launch table, Sonnet 5.5 posts 1,844 against 1,846 for the flagship Opus 5.5. Two points apart. Decrypt reports the model costs half what Opus 5.5 does.
Reading the table
Start with the jump. Sonnet 5 scored 1,449 on GDPval-AA, so one generation added 395 points. On Terminal-Bench 4.0 the gap is wider still: Sonnet 5.5 at 70.6%, Opus 5.5 at 66.4%, Sonnet 5 at 10.3%. That last figure is Anthropic’s own, and we’re reporting it as published.
Now look at where the mid-tier model still loses. FrontierCode 1.1: 46.2% against Opus 5.5’s 54.4%, an 8.2-point gap. CursorBench 4.0: 55.5% against 57.8%. OSWorld 2.1: 80.1% against 81.8%. Sonnet 5.5 ties or beats the flagship on terminal work and GDPval-AA, and trails it on the hardest coding and computer-use tests.
Put the two gaps side by side and the shape is clear. On GDPval-AA the price gap between the tiers is 50% and the score gap is 2 points. On FrontierCode the score gap is 8.2 points. A buyer running mostly scoped work sees a discount; a buyer running the hardest coding sees a trade-off. Which buyer a company is depends on what its people spend their week doing.
Anthropic’s own wording is careful. The model is strongest, it says, at well-scoped everyday tasks, fixing bugs, and building polished documents, slides and spreadsheets. That sentence draws two lines at once. Inside them sits work that has just gotten cheaper. Outside them sits work that still wants the flagship.
Cheaper depends on the dial
Anthropic says output is more than 30% faster than Sonnet 5 and per-task cost is up to 30% lower. Independent measurement tells a more mixed story. Artificial Analysis ranks Sonnet 5.5 second on its intelligence index at 56, two points behind Opus 5.5 at max effort. But at the highest reasoning setting the model averages about 193,000 output tokens per index task, the most Artificial Analysis has measured, which works out to roughly $7.60 a task, about 50% more than Sonnet 5.
These claims don’t contradict each other. They describe different settings. Turn the dial to max and the model thinks longer and the bill grows. Run it lower and the savings Anthropic describes show up. Customer numbers sit on that side: Balyasny Asset Management measured about 121,000 tokens per response against 497,000 on Sonnet 5, per MarkTechPost. Base44 ran 118 real app builds and got there in an average of 3.6 iterations, where Opus 5 took 7.7. Zendesk said tickets were processed 20% faster.
The same month’s other coding score
On September 13 we covered Real-SWE: the best coding agent cleared 38.8% of tasks pulled from real companies’ private codebases, where the reference fix touches a median of 11 files. Put that beside a 70.6% on Terminal-Bench and it reads like a contradiction. It isn’t. One test measures scoped terminal tasks, the other measures tangled legacy code. Anthropic drawing its own line at well-scoped work agrees with what Real-SWE found from the other direction.
What else shipped with it
The spec sheet is the ordinary kind. A 1M-token context window, 128K maximum output, adaptive thinking with five levels from low to max, and a June 2026 knowledge cutoff, according to Unite.AI’s summary and MarkTechPost. It is the first Sonnet to launch with the cyber safeguards built for Anthropic’s most capable models, with 99.43% recall on cyber-harm coverage. That last figure is Anthropic’s own measurement.
Customer quotes lean on quality more than speed. Epic Games said the model cleared the same bar as a higher tier on system-design audits, and Atlassian said customers can run its Rovo agents up to 30% faster than on Sonnet 5. Each of those is a vendor-supplied testimonial in a launch post, and worth weighting that way. Base44’s iteration count and Balyasny’s token count are the harder numbers, since both are measurable by anyone with the same workload.
The pattern to notice across all of them is the unit of measurement. None of them is “jobs replaced.” They’re minutes per ticket, iterations per build and tokens per answer. Those are the numbers finance teams convert into headcount plans, which is why a cheap, fast model on a familiar cloud matters to labor before it matters to any leaderboard.
What it means for the people doing that work
Start with who does documents, slides and spreadsheets all day. That’s the junior-analyst layer we wrote about when OpenAI shipped its financial-services product on September 10: first- and second-year staff building the model, the deck, the reconciliation. There’s now a cheaper tier aimed at the same output.
Second is the junior developer handling small bug tickets. Stanford’s AI Index has already shown the 22-to-25 software cohort down nearly 20% since 2024, before this model existed. Third is front-line support, which is what Zendesk’s 20% is measuring.
On timing, Sonnet 5.5 arrived on day one across the Claude Platform, AWS, Google Cloud and Microsoft Azure, so a company already buying through one of those doesn’t need a new procurement path. That matters more than the benchmark. Models that need a new contract move at the speed of legal review; models that appear in an existing cloud console move at the speed of a config change.
The forward question isn’t whether a 1,844 beats a 1,449. It’s which tasks a manager tags as scoped enough to hand to the cheaper tier at the next budget cycle, and how many entry-level requisitions quietly go unopened as a result.
What to watch
Three things. Independent cost per task at low and medium settings, to see whether Anthropic’s “up to 30% cheaper” holds outside its own tests. Whether the next round of private-codebase benchmarks like Real-SWE includes Sonnet 5.5, and whether that 38.8% moves. And how much of the customer language, tickets 20% faster, half the iterations, turns into fewer junior openings in Q4 hiring plans.
