Sonnet has always been first to ship models that actually get work done

For a lot of developers, the whole "AI that just does the work" wave started with Sonnet: Claude Sonnet 3.5, 3.6, and 3.7 were the first models to really turn heads with coding and tool use. Lately, though, the most visible capability gains have come from the pricier Opus line, and this Sonnet track fell behind. What Sonnet 5 is meant to do is close that gap back up.

Compared to the previous-generation Sonnet 4.6, Anthropic says Sonnet 5 makes clear gains across the key areas tied to agentic performance: reasoning, tool use, coding, and knowledge work.

How much intelligence the same dollar buys now

Anthropic released two cost-versus-performance curves comparing Sonnet 5, Sonnet 4.6, and Opus 4.8 across different effort tiers — the x-axis is cost per task, the y-axis is benchmark score. The takeaway: Sonnet 5 (orange line) beats Sonnet 4.6 (grey line) across the board, covers a far wider cost range than Opus 4.8 (yellow line), delivers a clear efficiency boost at mid tiers, and at its highest tier ties Opus 4.8 on some tasks.

Official cost-performance curves: the previous-generation Sonnet 4.6 clearly can't reach Opus 4.8; Sonnet 5 covers a far wider cost range and ties Opus 4.8 on some tasks at its top tier. Sonnet 5 in the chart is priced at the standard $3/$15 — at the limited-time $2/$10, actual cost is even lower. xhigh = the highest effort tier. Source: Anthropic's official site.

Spend more effort, think one step further: how one model pulls off both cheap and top-tier

The reason Sonnet 5 can span such a wide price range comes down to a mechanism called effort (a compute/reasoning-intensity tier): the same model, but you get to choose how hard it "thinks." Lower tiers are cheaper and faster but may be less thorough; higher tiers spend more compute on repeated reasoning and self-checking — more accurate answers, but pricier and slower.

Early user feedback: no nagging required — it checks its own work

Anthropic says feedback from early-access partners has been remarkably consistent: Sonnet 5 is noticeably more capable at working autonomously than previous generations. Here's what testers described in objective terms.

• On complex tasks, it keeps going until the job is done — earlier Sonnet generations often stopped partway through.
• It checks its own output for correctness without being explicitly asked to.
• The pricing for this kind of autonomous work is still quite appealing.

Safer overall — but cyberattack capability was deliberately held back

Pre-deployment safety evaluations show Sonnet 5 is safer overall than Sonnet 4.6: better at refusing malicious requests, more resistant to prompt injection (where an attacker hides malicious instructions inside a webpage or email the model is processing, trying to hijack the model into following the attacker's instructions instead of the user's), and less prone to hallucination and sycophancy. In an automated behavioral audit covering a range of problematic behaviors, it scored lower overall (meaning safer), though still higher than the stronger Opus 4.8 and Claude Mythos Preview.

Rate of problematic behavior across models in Anthropic's automated behavioral audit: Sonnet 5 is overall lower than Sonnet 4.6 (safer), but higher than Mythos Preview and Opus 4.8. Full list in system card section 6.4. Source: Anthropic's official site.

Cybersecurity is the one area deliberately held back. Anthropic says it did not specifically train Sonnet 5 on cybersecurity tasks: it can handle routine, benign network-related tasks, but on evaluations for developing potentially harmful software exploits, it performs notably worse than Opus 4.8 and Mythos 5.

A concrete test: can it hack its way through Firefox

"Weak at cyberattacks" sounds abstract, so Anthropic gave a concrete number: having each model develop an exploit for a vulnerability in the Firefox browser. This evaluation was jointly developed by Anthropic and Mozilla, and all vulnerabilities involved have already been fixed in Firefox 148.

Neither Sonnet model produced a single complete, working exploit (both at 0.0%). Sonnet 5 only edged out Sonnet 4.6 slightly on partial success rate, which Anthropic attributes mostly to spillover from general intelligence gains rather than targeted training. For comparison, both Opus 4.8 and Mythos 5 have far stronger cyberattack capabilities than either Sonnet model.

Firefox 147 exploit-development evaluation (jointly developed by Anthropic and Mozilla; the relevant vulnerabilities have been fixed in Firefox 148): for each model, the left bar is full-success rate for a working exploit, and the right bar is partial-success rate. Both Sonnet models scored 0.0% on full success, with Sonnet 5 slightly ahead of Sonnet 4.6 on partial success; both are far below Opus 4.8 and Mythos 5. See system card section 3.2.4 for details. Source: Anthropic's official site.

Because Sonnet 5 is slightly stronger than the previous generation on this kind of task, Anthropic has enabled real-time cybersecurity protections by default — able to detect and block dangerous network-related use in real time, at the same tier as Claude Opus 4.7 and 4.8. Anthropic judges Sonnet 5's overall cybersecurity risk to be low, so these protections are less restrictive than the ones on Fable 5 (which blocks a far broader range of cybersecurity-related tasks).

Looks like a price cut, but the ruler changed too

Sonnet 5 uses a new tokenizer. Before a model can process text, it first has to be split into individual tokens for computation and billing. With the new tokenizer, the same piece of text may get split into more tokens — roughly 1.0 to 1.35x, depending on content type. So while the per-token price has dropped, the same passage now consumes more tokens, meaning the actual per-unit cost hasn't dropped by as much as it looks.

Anthropic states outright: the limited-time pricing was set so that this upgrade works out to be close to cost-neutral. That's why "price cut" needs air quotes — you have to do the math with token counts included. This tokenizer adjustment follows the same approach used for Claude Opus 4.7.

Available now: rollout, pricing, and how to choose

Sonnet 5 is available today across every plan: it's the default model for Free and Pro plans, and Max, Team, and Enterprise users can access it too; it's also live in Claude Code and the Claude developer platform, where developers can call it via the Claude API as claude-sonnet-5.

To accommodate the higher token consumption that comes with higher effort tiers, Anthropic has already raised rate limits across Chat, Cowork, Claude Code, and the Claude developer platform — so you can pick the tier that fits each project.

How to choose

Sonnet 5 closes that gap: it's priced well below the flagship, yet official benchmarks show it tying Opus 4.8 on some tasks at higher tiers. Early user feedback is consistent too — it keeps working on complex tasks all the way through instead of stopping halfway, and it checks its own work without being asked.

How can the same model be both cheap and flagship-capable? The key is a dial called effort.

effort (the compute tier) lets you choose how hard it "thinks": at a low tier, it's cheap and fast but may not be thorough enough; at a high tier, it spends more compute on repeated reasoning and self-checking — more accurate, but pricier and slower. Turn the dial from the lowest tier, low, all the way to the highest, xhigh, and the same Sonnet 5 spans everything from a cheap tier to near-flagship.

Anthropic gives figures as "dollars per million tokens," which doesn't mean much to most people. It's more intuitive translated into what the same task actually costs — take a mid-sized task producing about 1 million output tokens and compare what each of the three options costs.