Claude Sonnet 5: Near-Opus Agent Performance at $2 per Million Tokens (For Now)

The Cheapest Serious Agent Model on the Market — Until August 31

Anthropic released Claude Sonnet 5 on June 30, 2026, and the headline is the price: introductory API rates of $2 per million input tokens and $10 per million output tokens through August 31, 2026, after which pricing settles at the standard Sonnet-tier $3/$15. For comparison, Anthropic’s top-end Opus 4.8 runs $5/$25 — so during the introductory window, Sonnet 5 costs less than half the flagship rate.

Positioning is unambiguous: this is a model built to make AI agents cheaper to run in production. Anthropic’s launch emphasizes agentic capability — planning, using tools like browsers and terminals, and executing multi-step workflows autonomously — which is exactly the workload pattern that racks up enterprise bills, because agents burn tokens across dozens of tool calls per task rather than one request-response round trip.

What’s Actually New

Sonnet 5 narrows the gap with Opus 4.8 substantially — launch coverage and Anthropic’s own framing put its coding and agentic performance close to the flagship at a fraction of the cost. The technical spec sheet backs the agent-first positioning:

• 1M-token context window with up to 128K output tokens per request — room for long agentic sessions without aggressive context management.
• Adaptive thinking on by default: the model decides when and how much to reason per request. The old fixed “thinking budget” parameter is gone entirely.
• The full effort range, including xhigh: Sonnet 5 is the first Sonnet-tier model to support the extra-high effort setting recommended for the hardest coding and agentic tasks — previously an Opus-only capability.
• High-resolution vision (up to 2576px on the long edge), which matters for computer-use agents reading real screenshots.
• A new tokenizer that produces roughly 30% more tokens for the same text than Sonnet 4.6 — more on why that matters below.

Sonnet 5 is also now the default model for free and Pro users of the Claude apps, with availability across Max, Team, and Enterprise plans — so the consumer surface and the API tier moved together.

The Tokenizer Fine Print

One practical caveat hides under the pricing headline: because Sonnet 5 uses a new tokenizer, the same text tokenizes to roughly 30% more tokens than on Sonnet 4.6. Per-token prices are unchanged at the sticker level, but per-request costs won’t map one-to-one from your old baselines. If you’re migrating a production workload, re-measure with the token-counting endpoint against the new model rather than reusing counts from Sonnet 4.6 — and give max_tokens limits extra headroom, since a value tuned for the old tokenizer can truncate equivalent output.

Even with the tokenizer shift, the introductory window is a genuine discount for agent workloads. Run your own numbers with our LLM API cost calculator — for a typical agent task mixing large inputs (context, tool results) with moderate outputs, the intro pricing lands meaningfully below anything else in the near-frontier class.

Where This Fits in the Market

The timing is hard to miss: Sonnet 5 arrived four days after OpenAI previewed its GPT-5.6 family under government-vetted limited access. While OpenAI’s release is capability-led and access-restricted, Anthropic’s is price-led and immediately available — a classic flanking move. For builders, the combined effect is the best buyer’s market for inference since early 2025: near-frontier capability is getting cheaper on a monthly cadence, and the introductory-pricing pattern (also used by other labs this year) rewards teams that can evaluate and switch quickly.

The economics of running agents have been the industry’s quiet bottleneck all year. Capability stopped being the limiting factor for many production use cases months ago; cost per completed task is what kills projects. A near-Opus model at $2/$10 changes those spreadsheets — a customer-support agent, a code-review bot, or a research assistant that was marginal at flagship prices may clear the bar comfortably now.

Getting Started

Access is self-serve: sign up for the Claude developer console, add billing, and generate a key — our updated guide on how to get access to the Claude API walks through the steps and current pricing for the full model lineup. The model ID is claude-sonnet-5.

Practical migration notes if you’re coming from Sonnet 4.6: omitting the thinking parameter now runs adaptive thinking (it previously ran without thinking), sampling parameters like temperature are no longer accepted, and instruction-following is more literal — prompts tuned to overcome older models’ quirks may now over-apply. Test before you flip production traffic, and use the introductory window as your evaluation budget.

Bottom line: Claude Sonnet 5 is a price-performance reset for agent builders. If your product runs multi-step AI workflows, the two months of introductory pricing are effectively a subsidized migration trial — worth taking.

Source: Anthropic announcement and launch coverage (June 30, 2026). Details as reported; verify specifics at the source.