- GPT-6.1 Terra: half the price, double the coding score (their numbers). Use.
- Gemini 4 Carbon: strongest Gemini yet, locked to cyber defenders. Watch.
- Genie 4: playable worlds, research preview. Ignore (for now).
- Mistral Large 4 weights and Kimi K3.1 Pro keep open models in the fight.
OpenAI’s Terra makes “good enough” a lot better
The most important model of the week isn’t the smartest one. It’s the one you can afford to run a million times.
On Oct. 26 OpenAI released GPT-6.1 Terra, its fast workhorse for computer use, visual reasoning, writing and coding. The headline: standard API prices are 50% lower than GPT-5.6 Terra and GPT-6.1 Sol (reported at $1 in / $5 out per million tokens).
By OpenAI’s own scoreboard:
- Terminal-Bench 4.0: 50.3% at max effort, up from 21.5% for GPT-5.6 Terra.
- Error rate on its tricky-prompt benchmark at low effort fell from 53.6% to 16.7%.
- OpenAI says it still recommends Sol and Astra for the gnarliest coding.
Why it matters: invoice entry, form filling and ticket triage were always “would automate if it were cheaper.” It’s cheaper.
Our verdict: Use. Re-run your most boring high-volume workflow on Terra and compare cost per task, not vibes.
Reach the people choosing their team's AI stack. One labeled sponsor per issue. Rate card →
Google’s best model has a velvet rope
Two weeks after Argon, Google announced Gemini 4 Carbon on Oct. 28, calling it its strongest model for cybersecurity and long-horizon tasks. It’s starting with a subset of cyber defenders, then trusted partners via its Fairwind Program. Developers come later, no date.
Pricing is already set at an introductory $2 in / $10 out per million tokens. Watch: great model, but a model you can’t call is a press release.
DeepMind’s Genie 4 generates a world (and the soundtrack)
Genie 4 (Oct. 27) generates interactive worlds in real time, with audio like music and dialogue made right alongside. It’s a research preview for trusted testers, and DeepMind’s own notes admit sessions last a few minutes and object permanence drifts (via The Model Press). Ignore for work. Absolutely watch the demo videos at lunch.
The cost-per-task test. Take 20 real examples of one repetitive task. Run them through your current model and through Terra. Score each output pass or fail and divide spend by passes. If Terra wins, you just found budget. If it loses, you just saved yourself a migration.
Sources & limits: Benchmarks belong to the vendor unless the story names another source. Email hello@loadedtokens.com and the correction goes on this page.