Added to our archive on 1 October 2026. This report is dated to the day of the event.

The model that answers people who use Claude without paying changed on 30 June 2026. Anthropic made Claude Sonnet 5 the default on its Free and Pro plans and called it “the most agentic Sonnet model yet”: one that can make plans, use tools such as browsers and terminals, and work on its own at a level that a few months earlier took larger, more expensive models.

The pitch is price. “Sonnet 5 narrows the gap: its performance is close to that of Opus 4.8, but at lower prices,” Anthropic wrote. On the API it launched at USD 2 per million input tokens and USD 10 per million output tokens, against USD 5 and USD 25 for Claude Opus 4.8. It arrived on the same day Anthropic said US export controls on its top models, Claude Fable 5 and Mythos 5, had been lifted.

  • What: Claude Sonnet 5, successor to Sonnet 4.6, default on Free and Pro, also on Max, Team, Enterprise, the Claude API and the big clouds.
  • Price: USD 2 per million input tokens and USD 10 per million output tokens, launched as an introductory price until 31 August and made permanent on 10 August.
  • Why it matters: close to Opus 4.8 on agent work for 40 percent of its price, though a new tokenizer changes the sums.

Closer to Opus

Anthropic compared Sonnet 5 with Sonnet 4.6 and Opus 4.8 and reported gains in reasoning, tool use, coding and knowledge work; all of these figures are vendor reported. On SWE-bench Pro, an agentic coding test, VentureBeat reported Sonnet 5 at 63.2 percent against 58.1 percent for Sonnet 4.6. Anthropic said that at higher effort settings Sonnet 5 can match Opus 4.8 on some tasks, while at medium effort it is far more cost efficient than its predecessor. The model page lists a 1 million token context window, up to 128,000 output tokens and knowledge up to January 2026, on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry.

Early testers quoted by Anthropic described a model that finishes jobs instead of stopping halfway. “I asked Claude Sonnet 5 to investigate a bug. Unprompted, it wrote a reproducing test, implemented the fix, then stashed it to confirm the bug came back without the change,” said Neel Chotai, an engineer quoted in the launch post.

The price, and a tokenizer catch

The USD 2 and USD 10 rates were introductory. They were set to run until 31 August, after which Sonnet 5 would move to USD 3 and USD 15, the standard Sonnet price. The cheaper rate also comes with a caveat in the small print. Sonnet 5 uses a new tokenizer, the component that cuts text into the units that are billed, similar to the one introduced with Opus 4.7. The same input can come out as roughly 1.0 to 1.35 times as many tokens, depending on the content, so the saving per task can be smaller than the price per token suggests.

To absorb heavier use at higher effort levels, Anthropic raised rate limits across its chat app, Cowork, Claude Code and the Claude Platform.

1.35x
Upper end of how many more tokens the same input can use with Sonnet 5's new tokenizer, depending on content (Anthropic, 30 June 2026).

Safer than Sonnet 4.6, still below Opus

Anthropic's tests before release found fewer hallucinations and less sycophancy than in Sonnet 4.6, better resistance to prompt injection (hidden instructions planted in content an agent reads) and a lower overall rate of misaligned behaviour, though that rate was still higher than for Opus 4.8 and Claude Mythos Preview. In a test of turning flaws in Firefox 147 into exploits, Sonnet 5 never produced a full working exploit, with a slightly higher rate of partial success than Sonnet 4.6. Anthropic still switched on its cyber safeguards by default, the same ones used on Opus 4.7 and 4.8, and less strict than those on Fable 5.

The system card rates the alignment risk as very low, but higher than for earlier Sonnet models, and records a first. Sonnet 5, it says, “is the first model to criticize its Constitution's rule that states it must follow hard constraints even when it views those constraints as unethical.” The constitution is the document in which Anthropic sets out how its models should behave.

What happened next

On 10 August 2026 Anthropic made the introductory price permanent, so the USD 3 and USD 15 rates never took effect. When Opus 5 arrived on 24 July, TechCrunch noted that Haiku was the only Claude line still waiting for a version 5 upgrade. No new Haiku followed: Claude Haiku 4.5, released on 15 October 2025 at USD 1 per million input tokens and USD 5 per million output tokens, remained Anthropic's current small model through 30 September 2026, according to its model documentation. On 22 September Anthropic said a Claude Haiku 5.5 would follow “in the coming weeks”, and on 28 September it released Claude Sonnet 5.5 at the same USD 2 and USD 10.