Anthropic quietly reset the price-to-performance curve for AI coding on June 30. Claude Sonnet 5, made the default model for every free and Pro user on July 1, scores 85.2% on SWE-bench Verified while charging $2 per million input tokens, which lands it within a point of the flagship Opus 4.8 at roughly a quarter of the cost. This is not a flashy new frontier record. It is something more disruptive: frontier-class coding at a price low enough that the expensive models now have to justify themselves.

  • Sonnet 5 hits 85.2% on SWE-bench Verified, versus about 86% for Opus 4.8 and 82.6% for GPT-5.5 on the same independent eval.
  • It is the new default on claude.ai for free and Pro users, so most people now get near-flagship coding at no cost.
  • Intro pricing is $2 in / $10 out per million tokens through August 31, then $3 / $15, still far below Opus 4.8's $5 / $25.
  • It ships with a 1M-token context window and runs on Anthropic's API, Amazon Bedrock, Google Vertex and Microsoft Azure.
Coding score versus price for leading models Sonnet 5 sits high on SWE-bench Verified at a low price, close to Opus 4.8 but far cheaper, while GPT-5.5 costs more for a lower score. SWE-BENCH VERIFIED (higher is better) PRICE PER 1M INPUT (higher is worse) → Opus 4.8~86% · $5 Sonnet 585.2% · $2 GPT-5.5 — 82.6% · $5 DeepSeek V4 Pro — 80.6% · $0.44 (open) Sonnet 4.6 — the model Sonnet 5 replaces Up and to the left is the sweet spot. Sonnet 5 is closest to it among closed models. genztech.blog
Fig 1 The trade every buyer actually cares about: coding accuracy against price. Sonnet 5 parks itself near Opus 4.8 on score but at a quarter of the token cost, while GPT-5.5 charges Opus money for a lower number.

What did Anthropic actually ship?

A mid-tier model that behaves like a flagship. Sonnet is Anthropic's workhorse line, sitting below the Opus flagship and above the tiny Haiku, and Sonnet 5 is the most agentic version of it yet. Anthropic tuned it for the thing developers now spend most of their time doing with these models: multi-step, tool-using coding runs where the model reads a repository, edits several files, runs tests and iterates. It ships with a 1M-token context window, enough to hold a substantial codebase in a single session, and it is available everywhere the enterprise buyers are: Anthropic's own API, Amazon Bedrock, Google Vertex AI and Microsoft Azure. The headline move, though, is putting it in front of ordinary users as the claude.ai default, so the model most people touch is now this one.

RelatedApple Tells 40 Ex-Staff at OpenAI to Preserve Data

Why does 85.2% on SWE-bench Verified matter?

Because SWE-bench Verified is the benchmark that stopped being easy to game. It measures the share of real, human-checked GitHub issues a model can resolve end to end, and scores have climbed from roughly 49% in late 2024 into the mid-80s in 2026. At 85.2%, Sonnet 5 is essentially tied with Opus 4.8's ~86% and clear of GPT-5.5's 82.6% on the same independent vals.ai evaluation. The practical read is that the gap between a mid-tier model and a flagship, measured on everyday coding, has narrowed to about a single point. For the vast majority of pull requests, refactors and bug fixes, that difference is invisible, and you are paying a large premium to chase it.

ModelSonnet 5Opus 4.8GPT-5.5Sonnet 4.6
SWE-bench Verified85.2%~86%82.6%lower
SWE-bench Pro63.2%69.2%58.6%lower
Input / output per 1M$2 / $10$5 / $25$5 / $30higher than 5
Context window1M1Mn/a200K
Free-tier defaultYesNo (Max)NoNo

How does the pricing actually work?

Aggressively, and on a clock. Through August 31, Sonnet 5 costs $2 per million input tokens and $10 per million output. After that it settles at $3 in and $15 out, still well under half of Opus 4.8's $5 and $25. Anthropic is doing two things at once here. The intro window pulls developers into building on Sonnet 5 now, while the token economics make it the obvious default for high-volume, always-on workloads like coding agents, batch code review and CI automation, where the bill scales with every call. When a model is both cheaper and nearly as capable, the burden of proof flips: you no longer justify using the cheap model, you justify reaching for the expensive one.

Who benefits, and who feels the squeeze?

Individual developers and small teams win outright, because the free claude.ai default is now a model that would have been a paid flagship a year ago. Startups building on the API win, because their per-token costs for agentic products just fell without a capability hit. The squeeze lands on the frontier tier itself. Opus 4.8 and GPT-5.5 now have to earn their premium on the genuinely hard 15% of tasks, the long-horizon refactors and security-sensitive work where the harder SWE-bench Pro still separates them (Opus 69.2% versus Sonnet's 63.2%). That is a real gap, but it applies to a shrinking slice of daily work. The rest of the market drifts toward the value model by default.

RelatedKimi K3 Is the Largest Open Model Ever, and It Is Not Cheap

SWE-bench Verified scores compared Sonnet 5 scores 85.2 percent, just below Opus 4.8 at about 86 percent and above GPT-5.5 at 82.6 percent and DeepSeek V4 Pro at 80.6 percent. Opus 4.8~86% Sonnet 585.2% GPT-5.582.6% DeepSeek V480.6% Independent vals.ai evaluation, real GitHub issues resolved end to end. genztech.blog
Fig 2 · benchmark On independent SWE-bench Verified numbers, Sonnet 5 (85.2%) sits a whisker below Opus 4.8 and comfortably ahead of GPT-5.5, at a fraction of either one's price. Full field on our AI coding leaderboard.

Where does Sonnet 5 sit in the lineup?

  1. Oct 2024Claude 3.5 Sonnet. First Sonnet to cross ~49% on SWE-bench Verified, kicking off the agentic era.
  2. 2025Sonnet 4.x line. Powers Claude Code and terminal agents; scores climb through the 50s and 60s.
  3. Jun 30 2026Sonnet 5 launches. 85.2% Verified at $2/1M; most agentic Sonnet yet, 1M context.
  4. Jul 1 2026Becomes the claude.ai default. Free and Pro users get near-flagship coding by default.
  5. Sep 1 2026Intro pricing ends. Rate settles to $3 in / $15 out per 1M.
What to watch · 2026
  • Does Opus keep its premium? Watch whether hard-task buyers stay on Opus 4.8 or drift to Sonnet 5 for cost.
  • OpenAI's response. GPT-5.5 now costs more for less on Verified. A price cut or a GPT-5.6 broad launch is the likely counter.
  • Agent economics. Cheaper tokens make always-on coding agents viable for more teams. Expect a wave of them.
  • Independent Pro scores. The mid-80s Verified plateau means SWE-bench Pro is now the real separator to track.

Our take

Sonnet 5 is the most important model release of the summer precisely because it is not trying to be the most impressive one. The frontier race for the top of SWE-bench Verified has quietly become a rounding-error contest, and Anthropic read that correctly: the leverage now is price, not another point of accuracy. By shipping near-Opus coding at a quarter of the cost and handing it to free users, it makes the default experience of Claude genuinely excellent and forces every rival to defend a premium that a shrinking share of work justifies. The flagships still matter for the hardest, longest, highest-stakes tasks, and the SWE-bench Pro gap is real. But for the daily reality of writing and fixing code, the value model just became the smart default, and that reframes the whole market around it.

Primary sources

Original analysis by GenZTech. Scores and pricing current as of July 2026. More at Anthropic.