The new mid-tier model is priced at $2 per million input tokens at launch, undercutting Opus 4.8 while closing the performance gap
Jaque Silva/NurPhoto via Getty Images
Anthropic launched Claude Sonnet 5 on Tuesday, a mid-tier model the company says can run autonomous AI agents at performance levels that previously required its larger and more expensive models. The model is now the default for Free and Pro plan users and is available across all subscription tiers, including Max, Team, and Enterprise.
Pricing starts at $2 per million input tokens and $10 per million output tokens, a rate Anthropic is holding through Aug. 31, 2026; after that date, the company said, token costs climb to $3 per million inputs and $15 per million outputs. At that price, Sonnet 5 undercuts Opus 4.8, though TechCrunch notes it does not go as low as Google $GOOGL's Gemini 3.5 Flash.
Anthropic said the model shows meaningful gains over its predecessor, Sonnet 4.6, on agentic tasks including reasoning, tool use, software coding, and knowledge work. Benchmark figures cited by TechCrunch put Sonnet 5 at 63.2% on an agentic coding test, placing it between Sonnet 4.6 at 58.1% and Opus 4.8 at 69.2%. For work requiring maximum accuracy, Anthropic said Opus 4.8 is still the stronger choice; developers running both models can tune how much effort each applies to manage the tradeoff between quality and cost.
Testers who tried the model before launch reported that it handled jobs that prior Sonnet releases abandoned midway and reviewed its own work for errors on its own initiative. "We handed Claude Sonnet 5 a two-part job — update Salesforce $CRM account tiers, send a launch announcement to enterprise contacts — and it finished end to end," Daniel Shepard, a senior engineer at Zapier, said in a statement. "That used to stall halfway."
Anthropic said safety testing found the new model less prone to problematic outputs than its predecessor, with improvements in refusing malicious prompts, withstanding prompt injection attempts, producing fewer hallucinations, and exhibiting less sycophantic behavior. Without dedicated cybersecurity training, the model scores considerably below Opus 4.8 on tasks that could enable harmful cyber activity, the company said, and Anthropic has activated protective cyber controls on it by default.
Cost pressures have made the timing relevant: Yahoo Finance reports that heavy AI token consumption grew unaffordable even for large technology firms, and Meta $META, Amazon $AMZN, and Uber $UBER are among those now actively curbing how much their systems spend on token processing.
The Sonnet 5 release follows a turbulent stretch for Anthropic. Federal authorities removed export restrictions on Anthropic's Fable 5 and Mythos 5 models on the same day, ending a shutdown of roughly 18 days that had been triggered by a government directive over cybersecurity concerns and cut off access for users everywhere. Separately, documents filed confidentially with the SEC earlier in June put the company's annualized revenue at $47 billion and its valuation at $965 billion as it moves toward a public offering.
Join 500,000+ readers who start their day with Quartz.
By subscribing, you agree to our Terms of Service and Privacy Policy.
