Anthropic released Claude Sonnet 5 today, calling it “the most agentic Sonnet model yet” and pricing it below the model it replaces. Through August 31 it runs at $2 per million input tokens and $10 per million output — a launch discount that settles to $3/$15 afterward, the same sticker as Sonnet 4.6. It’s live everywhere at once: Free, Pro, Max, Team, Enterprise, the API, and Claude Code.
The headline writes itself — cheaper model, better numbers, narrower distance to the flagship. The more useful read requires looking at what Anthropic did and didn’t put in the announcement.
What Anthropic Actually Claims
The framing is consistent and modest: Sonnet 5 is a “strict improvement over Sonnet 4.6” across reasoning, tool use, coding, and knowledge work, and it performs “close to that of Opus 4.8” while costing a fraction as much. That’s the same move Anthropic ran with Sonnet 4.6 in February, when the mid-tier model landed within a rounding error of the flagship on most benchmarks. The company is now repeating it deliberately as a product cadence: ship the flagship, then ship a Sonnet that eats most of its lunch at a third of the price.
What’s notable is what the launch page doesn’t hand you. The benchmark comparison against Sonnet 4.6 and Opus 4.8 is presented as a chart image, not transcribed scores — so the clean “Sonnet 5 hits X% on SWE-bench” line isn’t actually in the announcement. The third-party numbers already circulating (82.1%, 92.4%, take your pick) trace back to AI-generated aggregator posts, not Anthropic, and they don’t agree with each other. We’re not going to launder those into a citation. The honest statement is the one Anthropic makes itself: a strict step up over 4.6, approaching Opus 4.8, at lower cost. Treat anyone quoting a precise SWE-bench figure today with suspicion.
The Tokenizer Footnote That Isn’t a Footnote
Buried in the model’s documentation is a detail that matters more than the headline price: Sonnet 5 ships with an updated tokenizer that expands text into roughly 1.0 to 1.35× as many tokens as the previous generation for the same input.
Do the arithmetic before you celebrate the price cut. Input dropped from $3 to $2 per million tokens — a 33% reduction on paper. But if your workload sits at the high end of that tokenizer expansion, you’re now feeding 35% more tokens through the meter to process the same documents. The effective price per unit of work can land close to flat, or even up, depending on your content. Code and structured text tend to tokenize differently than prose, so the only real answer is to measure your own corpus before you assume the discount holds. This is exactly the kind of line item that quietly distorts an AI production cost model when nobody re-runs the numbers after a model swap.
None of this makes Sonnet 5 a bad deal. It makes the advertised price an incomplete one. The launch discount through August 31 is genuine; the standard pricing is where you should run your own tokens to see what you actually pay.
Safety Posture Is Part of the Pitch Now
Anthropic spends real space on what Sonnet 5 refuses to do. It’s better at declining malicious requests and resisting prompt-injection attacks, with lower hallucination and sycophancy rates than 4.6. And it launched with cyber safeguards enabled by default — Anthropic notes the model is a “substantially weaker” cyber operator than its Opus siblings and cannot develop full working exploits.
That last point is doing strategic work. The frontier labs have spent 2026 turning safety ceilings into product positioning — capability deliberately withheld or gated rather than maximized. We’ve watched this play out from Anthropic’s own Project Glasswing to OpenAI gating GPT-5.6 Sol behind government access. A mid-tier model that’s intentionally a weaker exploit-writer than the flagship is a feature in this environment, not a limitation. Anthropic is signaling that the cheap, widely-deployed model is also the safe-by-default one — which is the right model to have at the bottom of the funnel where volume lives.
Where This Lands for Teams
If you’re already on Sonnet 4.6, the upgrade math is easy: it’s a strict improvement, it’s drop-in compatible across the same surfaces, and the launch pricing is lower. Move, but re-baseline your token costs against the new tokenizer before you forecast savings.
If you’re paying Opus 4.8 pricing for production traffic, Sonnet 5 sharpens a decision you should already be making. The intelligent-routing argument — default to Sonnet, escalate to Opus only for tasks that demonstrably need it — gets stronger every time Anthropic ships a Sonnet that closes the gap. The “just use the best model” reflex was defensible when tier gaps were large. They aren’t anymore. Early partner feedback points at the obvious sweet spot: “Claude Sonnet 5 gives our agents a strong execution layer for multi-step software engineering work,” in the words of Anthropic technical staffer Zimu Li. Execution layer is the right phrase — Sonnet 5 is built to be the model that does the work in an agent loop, with Opus reserved for the planning and the genuinely hard reasoning.
What we’d caution against is treating the price drop as the whole story. The interesting number in this release isn’t on the benchmark chart Anthropic didn’t transcribe — it’s the tokenizer multiplier most teams won’t notice until the bill arrives. Run your own evals on your own data. The model is good. Whether it’s cheaper for you is a question only your workload can answer.