Claude Sonnet 5 Puts Opus Class Agents On A Mid Tier Budget

Anthropic released Claude Sonnet 5 on 30 June, describing it as its most agentic Sonnet to date and pricing it to make the flagship optional. The model plans, drives browsers and terminals and runs autonomously at a level that until recently required larger and more expensive systems. Claude Sonnet 5 became the default for Free and Pro users on claude.ai the day it launched, and is available across Max, Team, Enterprise, Claude Code and the API under the identifier claude-sonnet-5.

Pricing runs in two phases. Introductory rates of $2 per million input tokens and $10 per million output tokens apply through 31 August 2026, moving to $3 and $15 afterwards. Claude Opus 4.8 sits at $5 and $25. One detail belongs in any budget calculation: Claude Sonnet 5 uses an updated tokeniser, and the same input can map to between one and 1.35 times as many tokens depending on content type. A model that appears 60 per cent cheaper on the rate card may be closer to 40 per cent cheaper in practice.

The benchmark set shows a model that closes much of the gap to the flagship without quite reaching it. On SWE-bench Pro, Claude Sonnet 5 scores 63.2 per cent against Opus 4.8 on 69.2 per cent and Sonnet 4.6 on 58.1 per cent. On OSWorld-Verified computer use it reaches 81.2 per cent against Opus 4.8 on 83.4 per cent. On Humanity’s Last Exam with tools it scores 57.4 per cent, close to Opus 4.8 on 57.9 per cent. Two results break the pattern: on Terminal-Bench 2.1 it posts 80.4 per cent, above Opus 4.8, and on the GDPval-AA v2 knowledge work evaluation it records 1,618 Elo against Opus 4.8 on 1,615. The largest single improvement over its predecessor is a jump of more than twenty points on terminal work.

The context window is one million tokens with a maximum output of 128,000, extendable to 300,000 through a batch API beta header. Anthropic raised rate limits across Chat, Cowork, Claude Code and the platform to accommodate the higher token consumption that higher effort levels produce.

Safety documentation reports improvements across the board against Sonnet 4.6: better refusal of malicious requests, stronger resistance to prompt injection, and lower rates of hallucination and sycophancy. Real time cyber safeguards are enabled by default, the same ones used in Opus 4.7 and 4.8. Given the month that followed, in which prompt injection against coding agents became the sector’s most active attack surface, injection resistance in a model widely deployed inside agent harnesses is more than a line in a system card.

Reception split along predictable lines. Casual users found the value compelling. Power users running at high effort levels reported costs approaching Opus for output they judged slightly inferior, an outcome the tokeniser change makes easier to understand. Early independent testing placed it a little under Opus 4.6 at higher cost in some configurations, which is a reminder that headline rates and effective costs diverge once effort settings enter the picture.

The timing gave Claude Sonnet 5 an unusual role. It launched while Fable 5 and Mythos 5 were still suspended under the US export control directive, which made it briefly the most capable Claude model most people could reach, with Opus 4.8 the only higher option. Fable returned the following day. Sonnet 5 nonetheless became the workhorse for teams that concluded, correctly, that flagship access was going to remain unpredictable and expensive.

The strategic reading is that the mid tier is where the volume now sits. OpenAI made the same judgement with GPT-5.6 Terra at $2.50 and $15, offering GPT-5.5 class quality at half the cost. Grok 4.5 arrived at $2 and $6. Kimi K3 at $3 and $15. The two to three dollar input band has become the competitive centre of the market, and the flagships above it are increasingly bought for specific hard problems, with daily work handled below them.

For most agentic workloads Claude Sonnet 5 is the sensible default in the Anthropic range, with Fable reserved for work that justifies five times the rate.