Anthropic releases Claude Sonnet 5.5, narrowing the gap to Opus 5.5 on benchmarks at lower cost and faster output

Anthropic launched Claude Sonnet 5.5, the second release in its Claude 5.5 family, live immediately on AWS, Google Cloud and Azure under model ID claude-sonnet-5-5. Per-token pricing is unchanged from Sonnet 5 ($2 input / $10 output / $0.20 cache read), but Anthropic claims lower token usage cuts effective per-task cost by up to 30% and speeds output by over 30%. Anthropic's own benchmark table shows large gains over Sonnet 5 (e.g. Terminal-Bench 4.0 up from 10.3% to 70.6%) and positions Sonnet 5.5 close to Opus 5.5 on several tests. Anthropic also previewed a forthcoming Haiku 5.5 and added safeguards for high-risk cybersecurity requests.

Developers and enterprises get a concrete, immediately actionable option: same price, reportedly faster and cheaper per task, with a specific model ID to test today against existing Sonnet 5 workloads. If the efficiency claims hold up under independent use, this narrows the cost gap between mid-tier and flagship models, which matters for anyone budgeting large-scale API usage. The broader "closing in on Opus" and "cost war with GPT-6" framing is more about competitive positioning than a fundamental shift in what any single team can do differently right now.

All performance and cost figures come from Anthropic's own blog, relayed by The Decoder with no independent benchmarks or primary links; treat the 30% faster/cheaper claims as vendor marketing until third parties confirm them. The Terminal-Bench jump from 10.3% to 70.6% looks suspiciously large and likely reflects a harness or scoring artefact in the old Sonnet 5 baseline rather than a genuine capability leap. The unsourced claim that new cybersecurity safeguards close a "distillation attack" exploit allegedly used by Chinese labs is speculation dressed up as analysis, and the GPT-6 "cost war" framing is packaging rather than a demonstrated market shift.

1. 01Independent benchmark reruns (e.g. from third-party evals or Artificial Analysis) confirming or contradicting the Terminal-Bench and GDPval-AA gains within the next 2-4 weeks
2. 02Real developer cost reports comparing actual per-task spend on Sonnet 5.5 vs Sonnet 5 at scale, not just Anthropic's projected savings
3. 03Whether Haiku 5.5 ships on schedule and how it's priced relative to Sonnet 5.5, which will show if this is a coordinated tier refresh or a one-off
4. 04Any documented incident or lack thereof involving the new cybersecurity rerouting/distillation-attack safeguards that would validate or debunk the exploit-closing claim