On June 30, 2026, Anthropic released Claude Sonnet 5 (API model ID claude-sonnet-5), and it went generally available the same day: the default model for Free and Pro plans, available to Max, Team and Enterprise, and live in Claude Code and on the Claude Platform. Anthropic's pitch is blunt, "built to be the most agentic Sonnet model yet", a model that "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."
That is a big claim, and "most agentic" is marketing language, not a measured fact. So we did what we always do: we went to the primary sources, the official announcement, the model docs, the pricing pages and the system card, and split everything into verified facts, company-reported benchmarks, and our own analysis. The short version: the specs and pricing are rock solid, the benchmarks (all from Anthropic's own system card, with no independent third-party run published yet) put Sonnet 5 startlingly close to the much pricier Opus 4.8, and for anyone building agents or coding tools this is probably your new default model. Here are the numbers.
The verified spec sheet
First, the facts that are not in dispute, straight from Anthropic's model documentation. No invention, no rounding games.
| Claude Sonnet 5, verified specs | Detail |
|---|---|
| API model ID | claude-sonnet-5 (dateless pinned snapshot; AWS Bedrock anthropic.claude-sonnet-5, Google Cloud claude-sonnet-5) |
| Released | June 30, 2026, generally available across all plans and in Claude Code |
| Context window | 1M tokens |
| Max output | 128k tokens (up to 300k via the batch beta header) |
| Knowledge cutoff | January 2026 (reliable knowledge and training cutoff) |
| Input / output | Text and image input with vision, multilingual, text output |
| Thinking | Adaptive thinking (always available), effort defaults to high on the API and Claude Code |
| Latency | Fast (positioned as the best balance of speed and intelligence) |
The headline is the combination: a one-million-token context window and a January 2026 knowledge cutoff on a model Anthropic explicitly tunes for speed and agentic work. That is the workhorse profile, and it matters because agents burn tokens by the millions, so the model you run them on has to be both capable and affordable.
The price: cheap now, slightly less cheap in September
This is the part that reorders builders' defaults. Claude Sonnet 5's standard API pricing is $3 per million input tokens and $15 per million output tokens. But there is an introductory discount: $2 / $10 per million tokens through August 31, 2026, after which the standard $3 / $15 kicks in on September 1. The Batch API is cheaper still, $1 / $5 introductory.
To see why that is aggressive, line it up against the rest of the current Claude lineup. Sonnet 5 sits at a fraction of the Opus and Fable tiers while, as we will see, matching them on several evals.
Output price per million tokens, current Claude lineup (lower is cheaper)
Source: Anthropic models and pricing docs, June 2026 (verified). Output price per million tokens. One honest caveat: independent reports note Sonnet 5 uses a new tokenizer that produces somewhat more tokens per request, so the intro discount is closer to cost-neutral than to a true halving versus the previous Sonnet.
| Model | Price (in / out per MTok) | Context | Role |
|---|---|---|---|
| Haiku 4.5 | $1 / $5 | 200k | Fastest, near-frontier |
| Sonnet 5 | $3 / $15 ($2 / $10 to Aug 31) | 1M | Best speed + intelligence |
| Opus 4.8 | $5 / $25 | 1M | Most capable Opus-tier |
| Fable 5 | $10 / $50 | 1M | Most capable widely released |
The benchmarks: closing the gap to Opus
Important honesty note: every number below is from Anthropic's own system card. At the time of writing, on launch day, no independent third-party evaluation (Artificial Analysis and similar) had published its own run. Treat these as company-reported, directionally strong, but not yet independently confirmed.
The story the benchmarks tell is consistent: Sonnet 5 is a real generational jump over Sonnet 4.6, and on several evals it lands within a few points of Opus 4.8, a model that costs nearly twice as much on output. On the agentic coding benchmark SWE-bench Pro, the progression is clear.
SWE-bench Pro, agentic coding (company-reported, higher is better)
Source: Claude Sonnet 5 system card via The Decoder, June 2026 (company-reported). Sonnet 5 closes most of the gap to Opus 4.8 at a fraction of the output price.
It is even tighter on the broader, more expensive evals. Look at where Sonnet 5 essentially ties Opus 4.8 despite the price gap.
| Benchmark (company-reported) | Sonnet 5 | Sonnet 4.6 | Opus 4.8 |
|---|---|---|---|
| SWE-bench Pro (agentic coding) | 63.2% | 58.1% | 69.2% |
| Terminal-Bench 2.1 | 80.4% | 67.0% | not shown |
| OSWorld-Verified (computer use) | 81.2% | 78.5% | not shown |
| Humanity's Last Exam (with tools) | 57.4% | not shown | 57.9% |
| GDPval-AA v2 (knowledge work) | 1,618 pts | not shown | 1,615 pts |
Read those last two rows again. On Humanity's Last Exam with tools (57.4% vs 57.9%) and on the GDPval knowledge-work eval (1,618 vs 1,615 points), Sonnet 5 is statistically level with Opus 4.8, at $15 output versus Opus's $25. That is the entire thesis of this release in two numbers. The Terminal-Bench jump from 67% to 80.4% over a single Sonnet generation is the other eye-catcher: real progress on running things in a real shell, which is exactly what agents do.
Who is already using it
The announcement leans on named customers rather than anonymous praise, and the list is telling, because several of them are the AI-coding tools usually framed as a threat to incumbents. Anthropic quotes Replit ("handles sustained coding, tool use, and debugging well"), Cursor (excels on "brownfield code" with robust fixes), Lovable ("gets more done with less" and "refuses unsafe requests cleanly"), Salesforce (completed multi-step automations "end to end" that previously stalled), plus ClickHouse, Eve and Pace. When the companies building agentic products line up on day one, it is a signal the model is tuned for exactly that workload.
The catches to keep in view
It is a strong release, but a launch post will not lead with the asterisks.
- The benchmarks are Anthropic's own. Every figure here is from the system card. Independent evaluators had not published runs at launch. The numbers are likely directionally right, but wait for third-party confirmation before treating them as settled.
- The cheap price has a clock on it. The $2 / $10 introductory rate ends August 31, 2026. From September 1 it is $3 / $15, a 50% jump on both sides. Budget for the standard price, not the intro.
- A new tokenizer muddies the savings. Independent reports note Sonnet 5 tokenizes text into somewhat more tokens than the previous Sonnet, so for a given task the real cost is closer to flat than the headline discount suggests. Measure your own workload.
- "Most agentic Sonnet yet" and "close to Opus" are Anthropic's framing. The data supports a strong model, but the marketing superlatives are not independent verdicts.
Our read: is this the new default for builders?
What follows is our analysis.
For most coding and agentic workloads, yes, Sonnet 5 becomes the sensible default, and the reason is cost-performance, not raw peak intelligence. If a model lands a few points behind Opus 4.8 on the hardest coding eval but ties it on knowledge work and tool-use reasoning, while costing 60% of Opus on output (and far less during the intro window), then for the token-hungry reality of running agents the math points to Sonnet 5. You reserve Opus 4.8 and Fable 5 for the genuinely hardest reasoning, and you run the bulk of your agent traffic on Sonnet 5.
The deeper point, the one we keep making in this series, is that the model is a swappable component, not the product. The right architecture puts the model behind your own abstraction so you can move from Sonnet 4.6 to Sonnet 5, or to a cheaper open-weights model where quality allows, with a config change rather than a rebuild. That is how you capture a release like this on day one instead of in a quarter. It is exactly how we build AI features for clients: chosen per task on quality, cost and governance, over infrastructure you control (see our work). If you want help turning Claude Sonnet 5 into agents and product features that actually ship, and a stack that can swap the best model in as the next one lands, tell us about your project or get in touch, and we reply within 48 hours. For more on the fast-moving model landscape, read our breakdowns of GLM-5.2, the best open-weights LLM, SpaceX buying Cursor (a named Sonnet 5 customer) for 60 billion dollars, and Figma Motion against the AI-design wave.
Key numbers (as of June 30, 2026)
- June 30, 2026 Claude Sonnet 5 launches, generally available the same day, model ID
claude-sonnet-5. - $3 / $15 standard price per million tokens, with $2 / $10 introductory pricing through August 31, 2026.
- 1M tokens context window, 128k max output, January 2026 knowledge cutoff.
- 63.2% SWE-bench Pro (company-reported), versus 58.1% for Sonnet 4.6 and 69.2% for Opus 4.8.
- 57.4% vs 57.9% Sonnet 5 versus Opus 4.8 on Humanity's Last Exam with tools, effectively tied.
- 80.4% Terminal-Bench 2.1, up from 67.0% on Sonnet 4.6.
