Claude Sonnet 5.5: same $2/$10, ~30% fewer tokens per task
Anthropic's new default beats GPT-6 Sol on real-work evals and leaps to 70.6% on Terminal-Bench, while Sonnet 5 stays put. Plus: GLM-5.3-Flash now runs locally.

Copy markdown
Same sticker price, up to 30% cheaper in practice
Sonnet 5.5 keeps Sonnet 5's $2/$10 per-million-token pricing, but Anthropic says it finishes the same task with far fewer tokens — cutting real-world cost up to 30% and generating output about 30% faster. Point your Claude Code or API calls at it and your existing agent bills drop with no other changes.
Terminal-Bench jumps to 70.6%
It scores 70.6% on Terminal-Bench 4.0 at max effort, versus Sonnet 5's ~43% at the API default (10.3% at max) — a big step up in agentic command-line reliability. It's also the first Sonnet to finish Pokémon Red from screenshots alone.
Wins the real-work evals, trails on raw code
Sonnet 5.5 edges GPT-6 Sol on GDPval-AA and AA-Briefcase (real-occupation tasks) and lands within about 2 points of Opus 5.5 on GDPval-AA, but GPT-6 Sol still leads on FrontierCode. Make it your default for agentic and real-world work; keep Opus or Sol for the hardest pure coding.
1M context, cheap cache, and Sonnet 5 stays
It holds the 1M-token input window (128K output, 300K on the Batch beta), with cache reads at $0.20/M and batch at $1/$5. Sonnet 5 remains available at unchanged prices, so there's no forced migration — run both side by side before you switch.
GLM-5.3-Flash now runs on your own box
Zhipu's 320B/18B-active multimodal MoE now has llama.cpp support and GGUF quants from Q8 down to IQ2. Pull it with `llama-server -hf AtomicChat/GLM-5.3-Flash-GGUF:UD-Q4_K_XL --jinja` — a frontier-class open MoE you can self-host, with Q4_K_M the recommended default and IQ2 for tight VRAM.