Model Drop: Claude Sonnet 5
Claude's long awaited upgrade to its cheaper workhouse Sonnet line
Our last two drops were models you couldn’t touch (GPT-5.6 behind a US-government gate, Fable 5 yanked by an export-control suspension). Claude Sonnet 5 is the opposite, available everywhere now, and it’s the first Sonnet pitched as a near-Opus model at Sonnet money.
Model: Claude Sonnet 5 (claude-sonnet-5 on the Claude API, anthropic.claude-sonnet-5 on Bedrock, claude-sonnet-5 on Vertex AI). Dateless pinned snapshot like the rest of the 4.6-and-later generation.
Model type: Text + image input, text output. Vision and multilingual. No native image, audio, or video output.
Ship date: June 30, 2026
Maker: Anthropic (San Francisco)
Pricing: Introductory $2 / $10 per million input / output tokens through August 31, 2026, then $3 / $15 (the same list price Sonnet 4.6 carried). Standard Anthropic prompt caching (cache reads at 0.1x input) and a 50% Batch API discount apply. Sonnet 5 uses the tokenizer Anthropic introduced with Opus 4.7, so the same text runs roughly 1.0-1.35x more tokens than older Sonnet versions depending on content type. Still cheaper per token than Opus 4.8 ($5 / $25), GPT-5.5 ($5 / $30), and Gemini 3.1 Pro, and more expensive than Gemini 3.5 Flash.
Available on: claude.ai and the Claude apps (default model for Free and Pro, available to Max, Team, and Enterprise), Claude Code, the Claude Developer Platform, Claude Platform on AWS and Amazon Bedrock, and Microsoft Foundry. Google Vertex AI is listed as coming soon.
Headline benchmarks: Agentic coding 63.2%, against Opus 4.8 at 69.2% and Sonnet 4.6 at 58.1% (same axis Anthropic now leads its coding charts with). Anthropic says Sonnet 5 slightly outperforms Opus 4.8 on knowledge work. It charted Sonnet 5 against Sonnet 4.6 and Opus 4.8 across effort levels on BrowseComp (agentic search) and OSWorld-Verified (computer use), where Sonnet 4.6’s baselines sit at 78.5% OSWorld-Verified and 34.6% / 46.8% on Humanity’s Last Exam (no tools / with tools). On the launch’s cyber eval, building a working exploit for a real vulnerability in Firefox 147, Sonnet 5 never produced a full working exploit (0%), landing well below Opus 4.8 by design.
Other info: 1M-token context window (~555k words), 128K max output (up to 300K through the Batch API extended-output beta). Reliable knowledge cutoff and training cutoff both January 2026. Released under Anthropic’s Responsible Scaling Policy with the cyber safeguards from Opus 4.7 and 4.8 enabled by default. Anthropic reports a lower rate of undesirable behaviors (cooperation with misuse, deception) than Sonnet 4.6. System card published at launch.
More details: Introducing Claude Sonnet 5 and the Claude Sonnet 5 System Card.
What shipped
Anthropic released Claude Sonnet 5 today and made it the default model for Free and Pro accounts the same day, with Max, Team, and Enterprise able to select it and the API live across Anthropic’s platform, AWS, and Microsoft Foundry. The pitch: agentic work that used to demand the Opus tier now runs on Sonnet pricing. Anthropic calls it “the most agentic Sonnet model yet,” built to make plans, drive browsers and terminals, and run autonomously through long tasks where earlier Sonnets would stop short. The number on the box is the introductory rate, $2 in and $10 out per million tokens through August 31, which undercuts Opus 4.8, GPT-5.5, and Gemini 3.1 Pro.
The supporting evidence is real, but its narrower than the framing suggests. On Anthropic’s agentic-coding eval Sonnet 5 posts 63.2%, a solid jump over Sonnet 4.6’s 58.1% and within six points of Opus 4.8’s 69.2%, and Anthropic claims it edges Opus 4.8 on knowledge work. Sonnet 5 logs fewer undesirable behaviors than its predecessor (less cooperation with misuse, less deception), and on the Firefox 147 exploit-development test it never built a working exploit, which puts it well under Opus on the one capability Anthropic most wants capped at this price point.
Opus 4.8 still wins raw coding capability, the introductory price expires in two months, and the Opus 4.7 tokenizer means the same prompt bills 1.0-1.35x more tokens than it would have on Sonnet 4.6, so the per-task discount is smaller than the per-token sticker implies.
What’s new
Sonnet 5 is a real step over Sonnet 4.6, not a checkpoint refresh.
Opus-class agentics at Sonnet money. The whole release is built around closing the gap to Opus 4.8: 63.2% agentic coding against Opus 4.8’s 69.2%, a claimed knowledge-work edge over Opus, and autonomous browser and terminal runs Anthropic says used to require a larger model. No prior Sonnet has been positioned as a direct Opus substitute for agent loops.
Task completion, not task attempts. Anthropic’s central claim is that Sonnet 5 “finishes complex tasks where previous Sonnet models would stop short.”
A capability ceiling on purpose. Sonnet 5 ships with cyber safeguards on by default and was tuned to stay materially weaker than Opus at offensive security. It never developed a full working Firefox exploit in Anthropic’s eval. Shipping a flagship Sonnet with an explicit “this far and no further” on a dangerous capability is new for the tier, and it’s a deliberate product decision, not a benchmark miss.
Fewer bad behaviors than its predecessor. Anthropic reports Sonnet 5 cooperates with misuse and deceives less often than Sonnet 4.6. For anyone wiring this into a customer-facing agent, the alignment delta is the part that doesn’t show up on a coding leaderboard but shows up in production.
How and where to use it
Where it runs, what it’s actually for, and where you’ll want a different model.
Where it’s available
Claude.ai and the Claude apps, where it’s the default for Free and Pro and selectable for Max, Team, and Enterprise
Claude Code for the coding agent, with
effortdefaulting to highThe Claude Developer Platform for the API, plus Claude Platform on AWS, Amazon Bedrock, and Microsoft Foundry. Google Vertex AI is coming soon
What it’s good at
High-volume agentic work where you were paying Opus rates and didn’t need Opus capability
Long-horizon coding and refactors, browser and terminal automation, deep-research and search loops (BrowseComp, OSWorld-Verified), and knowledge work where Anthropic says it matches or beats Opus 4.8
What it’s bad at / shouldn’t be used for
The hardest coding, where Opus 4.8 still leads agentic coding by six points and you should keep paying for it
Offensive-security and exploit-development work, which Anthropic deliberately capped (use the right tool, and that tool isn’t a discounted Sonnet)
Cost-floor workloads where Gemini 3.5 Flash or Haiku 4.5 is cheaper and the extra capability is wasted. And any budget math that assumes the $2 / $10 rate is permanent, because it reverts to $3 / $15 on September 1 and the tokenizer is already charging you more tokens than Sonnet 4.6 did.
First impressions
Impressions are scarce, and these are write-ups, so take with a grain of salt.
The positives
TechCrunch framed the launch as a cost-and-reliability play and led with Anthropic’s own positioning line:
It can make plans, use tools like browsers and terminals, and run autonomously at a level that, just a few months ago, required larger and more expensive models.
The interesting tell is what Anthropic chose to brag about. Not a new capability ceiling, but doing the existing job cheaper and finishing it. When the headline is “you can stop paying for the big model,” the frontier has moved from “can it do this” to “what does it cost to do this at scale.”
The New Stack put the capability story plainly in its headline: Sonnet 5 “closes the gap with Opus 4.8, and is cheap until August.” A Sonnet landing within six points of the flagship on agentic coding, and reportedly ahead on knowledge work, is the first time the mid-tier has been a credible Opus substitute for real agent loops rather than a fallback you tolerate.
VentureBeat reported that Anthropic says Sonnet 5 “slightly outperforms Opus 4.8” on knowledge work. If that holds up in independent testing, it scrambles the default-model decision for any team that’s been routing knowledge tasks to Opus out of habit.
The negatives
The New Stack‘s own framing is also the sharpest critique: “cheap until August.” The $2 / $10 rate is a two-month promotion. It reverts to $3 / $15 on September 1, and because Sonnet 5 runs the Opus 4.7 tokenizer, the same workload bills 1.0-1.35x more tokens than it did on Sonnet 4.6. The real cost increase against last-gen Sonnet is hidden inside the token count, where most people won’t look until the invoice arrives.
VentureBeat summed up the positioning as a concession dressed as a win:
cost-efficiency and reliability, rather than raw capability, now differentiate AI models.
Translated: Sonnet 5 is not the most capable model Anthropic sells, and the company is leaning on price because Opus 4.8 still owns the capability lead. On the hardest agentic coding, Opus is six points ahead, and “close to Opus” is doing a lot of work in a sentence that also wants you to switch off Opus.
The same VentureBeat piece flags the timing: the discount lands as Anthropic “races toward a blockbuster IPO.” An aggressive introductory price that expires right after the quarter is a customer-acquisition wedge with a roadshow attached. It tells you the usage numbers matter to the company right now in a way that has nothing to do with whether the price is sustainable for you in September.
Jake’s take
Sonnet 5 appears to close most of the gap to Opus 4.8 around use cases like “run this agent a few hundred times a day and don’t let it stall halfway,” while billing at a fraction of the rate. A cheap, reliable, tool-using default is precisely the model you want underneath internal automation that has to run constantly and can’t justify Opus economics. Take whatever you’ve got pointed at Opus 4.8 for agentic and knowledge work, swap the model string to claude-sonnet-5, and watch the completion rate and the bill. If Anthropic’s knowledge-work claim survives contact with your real prompts, you can cut your inference cost by more than half on a chunk of your stack.
But the pricing theater sucks, as usual. The $2 / $10 rate is a two-month coupon that expires September 1, and the Opus 4.7 tokenizer means even the coupon buys fewer effective tokens than Sonnet 4.6 did, so the “cheaper Sonnet” is quietly more expensive per task than the previous Sonnet once the promo ends.
With the IPO coming, its likely that Anthropic wants to juice the usage numbers before the roadshow, then reset to $3 / $15 and let switching costs do the rest. None of that makes Sonnet 5 a bad model. It just makes the headline price a marketing number, and if you build your unit economics on the June rate you’re going to have a bad September. The capability story is genuinely good and the cyber ceiling is the kind of restraint I like seeing personally. But don’t let “default model, basically free, almost as good as Opus” do your math for you.