Claude Sonnet 5 launched on June 30, the same day Anthropic lifted the export controls on Fable 5, and it got a fraction of the attention. That's a shame, because for most day-to-day engineering work, Sonnet 5 is the more relevant release of the two.
Where it sits now
Sonnet models have historically been the "good enough, much cheaper" option compared to Opus. Anthropic's own framing of Sonnet 5 is more confident than that: it's described as narrowing the gap with Opus 4.8, close enough on several agentic benchmarks that the choice between them becomes a genuine cost-versus-performance decision rather than a capability gap you're stuck accepting.
Concretely, Sonnet 5 is now the default model on Free and Pro plans, and it's available across Max, Team, Enterprise, Claude Code, and the Claude Platform via the API as claude-sonnet-5.
Pricing, and a tokenizer catch worth knowing about
Sonnet 5 launched at an introductory $2 per million input tokens and $10 per million output tokens through August 31, 2026, moving to standard pricing of $3 and $15 per million after that. Opus 4.8, for comparison, runs $5 and $25.
There's a detail buried in the footnotes that's worth flagging if you're budgeting API usage: Sonnet 5 uses an updated tokenizer, the same kind of change Anthropic made with Opus 4.7. The same input text can map to somewhere between 1.0x and 1.35x more tokens than it did under the old tokenizer, depending on content type. Anthropic says the introductory pricing was set specifically to make the transition roughly cost-neutral, but if you're tracking token usage dashboards for cost forecasting, don't be surprised if the raw token count per request ticks up even though your bill doesn't move much.
What actually improved
Early access partners kept describing the same thing in different words: Sonnet 5 finishes tasks that Sonnet 4.6 used to stop halfway through, and it checks its own output without being explicitly told to. One partner had it update Salesforce account tiers and send a launch announcement end to end in a single pass. Another described handing it dozens of real pull requests and getting tested, verified results back without hand-holding. That's a meaningfully different agentic profile than a model you have to babysit through each step.
On the safety side, Anthropic's pre-deployment testing found lower rates of hallucination and sycophancy compared to Sonnet 4.6, along with better resistance to prompt injection and hijack attempts during agentic tasks. It does show a somewhat higher rate of misaligned behavior than Opus 4.8 or the Mythos preview on Anthropic's internal behavioral audit, so it's an improvement over its own predecessor without matching the safety profile of the more expensive tier above it.
Should you switch?
If your workloads are mid-complexity agentic tasks — sustained coding sessions, tool use, multi-step automation — Sonnet 5 is worth benchmarking against whatever you're running today before you assume you need Opus-tier pricing. Reserve Opus 4.8 or Fable 5 for the genuinely hardest, longest-running work, and let Sonnet 5 carry everything in between.
