Anthropic Launches Claude Sonnet 5, Targeting Low-Cost AI Agents
Alina Collins
Anthropic released its mid-tier Claude Sonnet 5 on June 30, scoring 63.2% on agentic coding benchmarks at a price below flagship Opus 4.8 and rival GPT-5.5 — making it the new default for free and Pro users as mid-tier AI competition shifts from 'who can do agents' to 'who does it cheapest with the least hand-holding.'
How strong is Sonnet 5, exactly?
Agentic coding benchmark: 63.2%, up from Sonnet 4.6's 58.1% but below flagship Opus 4.8's 69.2%.
On knowledge-work benchmarks, Sonnet 5 slightly outperformed Opus 4.8. This means → it is not uniformly weaker than the flagship — strengths depend on task type.
In plain terms = for writing, analysis, and research tasks, Sonnet 5 is already as good or better; for high-precision complex coding, Opus 4.8 still leads.
How much cheaper — and for whom?
Launch pricing through August 31: $2 per million input tokens, $10 output. After August 31, input rises to $3; output stays at $10.
Anthropic says this undercuts Opus 4.8, OpenAI's GPT-5.5, and Google's Gemini 3.1 Pro — but remains above Google's Gemini 3.5 Flash.
This means → Sonnet 5 targets enterprise developers who want agentic capability but balk at flagship pricing — roughly 90% of the performance at roughly 60% of the cost.
What does "agentic capability" actually mean?
Agentic capability means a model can plan autonomously, use tools like browsers and terminals, and complete multi-step tasks end-to-end — not just answer individual questions.
A Zapier engineer tested it live: Sonnet 5 was given a two-step task — update a Salesforce account hierarchy, then send an onboarding announcement to enterprise contacts. It completed the whole sequence. Previous models would stall midway.
This reflects a rapid convergence: mid-tier models are closing the reliability gap with flagships. OpenAI's GPT-5.6 Sol and Google's Gemini 3.5 Flash are making the same pitch.
How safe is it — and where are the gaps?
Sonnet 5 outperforms Sonnet 4.6 at rejecting malicious requests and resisting prompt injection — a technique where disguised instructions trick a model into executing harmful actions. Hallucination and sycophancy rates are also lower.
But Anthropic acknowledges alignment performance still trails Opus 4.8 and Claude Mythos Preview, and its capability for dangerous cybersecurity tasks is "far below" the Opus line.
In plain terms = Anthropic is proactively saying "this model can't do dangerous things." That reads less like a technical disclosure and more like a safety card handed to regulators.
Why stress "it can't do dangerous things"?
Axios reports that Sonnet 5's launch coincides with ongoing talks between Anthropic and the Trump administration over its models — talks that specifically cover Sonnet 5's release.
The more powerful Mythos model is currently available only in limited access. Fable 5 was previously pulled at government request on safety grounds and is expected to return soon.
This means → Anthropic's product cadence is no longer purely technology-driven — government approval is becoming a hidden variable on every AI company's product calendar.
What does this mean for enterprise buyers?
Sonnet 5 is now fully available to Max, Team, and Enterprise subscribers, and is the default model for free and Pro tiers.
Competition is shifting from "who can build agents" to "who delivers them at the lowest cost with the least human intervention."
This reflects a change in how enterprises evaluate AI models: the question is no longer "which is strongest?" but "which offers the best performance per dollar while saving the most labor?"
Content is for reference only, not financial advice.