Claude Opus 5.5 fast mode drops the beta header: 2.5x output
One request flag streams Opus 5.5 at 2.5x — plus $2,599 RTX Spark laptops, a free DeepSeek coder on OpenCode, and FTC subpoenas for the big two.
Copy markdown
Flip one flag, stream 2.5x faster
Opus 5.5 (`claude-opus-5-5`) fast mode no longer needs the `fast-mode-2026-02-01` beta header or an access request — just set `speed:"fast"`. You stream up to 2.5x the output tokens/sec (it speeds generation, not time-to-first-token); Opus 5.5 lists at $8/M in and $40/M out, with fast mode billed as a multiplier on top and its own rate limit across the Start, Build and Scale tiers.
OpenCode hands you a free DeepSeek coder
OpenCode's Zen gateway now serves DeepSeek V4.1 Flash at no cost with a 200K-token context (1M on the paid tier). If you keep hitting limits on paid coding models, that's a zero-dollar agent backend you can point your terminal at today.
RTX Spark laptops open to order at $2,599
NVIDIA's Grace-Blackwell RTX Spark laptops go up for preorder from Surface, ASUS, Dell, HP, Lenovo and MSI, running ~$2,599 (Surface Laptop Ultra) to $3,799 (Dell XPS 16). The builder pitch: enough on-device memory to run this week's big open weights — Muse Spark, Nemotron — off your lap instead of a rented GPU.
FTC subpoenas OpenAI and Anthropic
The FTC has reportedly issued ~50-page civil investigative demands to both labs over security incidents and safeguards, escalating the probe it opened in late September. Nothing changes in your API today, but formal discovery into the two vendors most builders lean on is a dependency risk worth tracking.
Accenture caps Claude at $29 a head
Accenture now limits central Claude funding to $29 per employee per month after some staff reportedly ran $500–$1,000 bills — joining Uber, Walmart and JPMorgan in clamping down on AI spend. Build on a corporate plan and you should expect tighter token budgets, so wire in cheaper model fallbacks now.
Research radar: Claude agents fine-tuned a model solo
In the new Athanor paper, four Claude Opus 5.5 agents ran a 104-hour autonomous post-training campaign that lifted a 4B base model's dev score from 15.9 to 67.3 — matching Qwen3-4B-Instruct. An early look at agents owning a fine-tuning run end to end.