Four Releases, One Theme: Cost Compression

Four notable launches shipped between October 26 and 31. The week’s headline Terminal-Bench number belongs to Kimi K3.1 Pro at 67% — an independent Artificial Analysis figure that edges out Claude Opus 5.5’s 66.4% in Moonshot’s own comparison table. But the louder signal is price: GPT-6.1 Terra at $1/$5 per million tokens is half the price of GPT-6.1 Sol and one-tenth the price of GPT-6 Astra per output token, according to DataCamp’s coverage of OpenAI’s announcement.

The output token gap between the cheapest new model this week (Mistral Large 4 API at $2.10/M) and GPT-Rosalind-5.5’s reference price ($25/M) is more than 11×. That spread reflects differentiation strategy — specialist access control at the top end, commodity inference at the bottom — more than raw capability.

Rankings: Week of 27 October 2026

*Vendor-reported benchmarks. †Artificial Analysis independent evaluation. Best public Terminal-Bench 4.0 score: Claude Sonnet 5.5 at 70.6% (benchlm.ai). GPT-6 Astra: 58.2% per Moonshot’s comparison table.

Cost Per Task: The Number That Determines Production Budget

Terminal-Bench tells you what a model can do. Cost per task tells you what it will cost when it runs. Assume a mid-size agentic coding task consumes 50,000 input tokens and 10,000 output tokens — roughly the load for a full-context codebase analysis with a structured response. All prices from vendor pricing pages as of October 31, 2026. No independent verification of actual token consumption for this task definition.

Prices: Mistral from Mistral API docs (via cleus.ai); GPT-6.1 Terra from DataCamp/OpenAI announcement; Kimi K3.1 Pro from aicybr.com. GPT-Rosalind-6 pricing not published at launch.

At 10,000 tasks per month, Kimi K3.1 Pro Standard costs $1,440 more than Terra. Whether its 15–16-percentage-point Terminal-Bench lead translates to that value in your pipeline depends on whether your production tasks actually resemble Terminal-Bench problems — which is a measurement you need to make yourself, not accept from a leaderboard.

Model Notes

1 — GPT-6.1 Terra (OpenAI, Oct 26)

Terra is OpenAI’s efficient tier for late 2026: $1 input, $5 output, $0.05 cached input per million tokens, and a 1.05-million-token context window. OpenAI claims it completes Agents’ Last Exam tasks over three times faster than GPT-6 Astra or GPT-6.1 Sol (vendor-reported). Terminal-Bench 4.0 at max effort sits at 50.3% — up 28.8 percentage points from GPT-5.6 Terra’s 21.5%. Those figures are OpenAI’s own. The one independent data point in the agentic category is MazeBench at 16%, compared with GPT-6 Astra’s 23% (benchlm.ai). The multimodal story is stronger: MMMU-Pro at 85.5% is tied for the top verified result, and FrontierMath v2 Tiers 1–3 at 90.877% comes from Epoch AI independently. The DataCamp analysis notes that OpenAI’s claim of “roughly 9× lower cost than Claude Opus 5.5 on computer use” is described as selective. Full benchmark breakdown at benchlm.ai.

2 — Kimi K3.1 Pro (Moonshot AI, Oct 23)

The week’s highest public Terminal-Bench 4.0 score: 65.7% by Moonshot’s own figures, 67% when Artificial Analysis ran it independently. Moonshot’s own comparison table lists Claude Opus 5.5 at 66.4% and GPT-6 Astra at 58.2%, so treat the 67% figure as the one to compare — but note that benchlm.ai shows Claude Sonnet 5.5 at 70.6% as the best verified score, which is a different Anthropic model. Context window is confirmed at 1,048,576 tokens. Standard pricing is $2 input / $10 output per million; Highspeed doubles those rates and claims 2–3× the speed; cache hits cost $0.20 per million on both tiers. Open weights are scheduled for November 4, 2026; licence terms and minimum hardware for self-hosting have not been published. Full pricing and context-window details at aicybr.com.

3 — Mistral Large 4 weights (Mistral AI, Oct 29)

The weights landed October 29. The modified MIT licence permits self-hosting and fine-tuning; reselling the model as a hosted service requires a separate agreement from Mistral. Mistral reports approximately 1 trillion total parameters with roughly 52 billion active per call. A quantised version runs on four Nvidia B200 or B300 GPUs. Terminal-Bench 4.0: 41% (Mistral-reported). Mistral also claims 47% fewer tokens than GLM-5.3 on Artificial Analysis’s Coding Agent Index when handling the same tasks. API pricing stays at $0.70 input / $2.10 output per million tokens. Our full EU self-hosting cost breakdown — GPU sizing, power cost, and breakeven versus the API — is at vortx.ch/mistral-large-4-weights-drop-the-eu-self-hosting-math/.

4 — GPT-Rosalind-6 (OpenAI, Oct 30)

GPT-Rosalind-6 reuses GPT-6 Astra weights with relaxed refusals on dual-use biological queries, under a trusted-access programme that requires business verification and, for government applicants, a separate biosecurity pathway. Benchmark gains over GPT-Rosalind-5.5 (all OpenAI-reported): MedChemBench v7 rises from 28.9% to 49.6% (99 problems), GeneBench Pro v13 from 16.7% to 37.7% (129 problems), LifeSciBench Gold from 44.0% to 61.3% (94 problems). OpenAI rates it High on biological and chemical capability, below Critical. Pricing was not published at launch; GPT-Rosalind-5.5 was priced at $5 input / $25 output per million tokens. Not relevant to general software engineering teams. Coverage at AI Weekly.

For Swiss & EU teams

Mistral Large 4’s weight release is the most directly actionable item this week for EU and Swiss engineering teams. The modified MIT licence permits deployment in Swiss and EU data centres without routing inference traffic through Mistral’s API — directly relevant for teams managing data under FADP or EU AI Act requirements. Our cost breakdown covering GPU hardware sizing, power estimates, and the API breakeven point is at vortx.ch/mistral-large-4-weights-drop-the-eu-self-hosting-math/. For GPT-6.1 Terra: OpenAI has not announced EU regional processing endpoints for this model as of October 31, 2026.

What to Watch Next Week

Two unresolved items carry into November. First, Kimi K3.1 Pro’s open weights (due November 4) will determine whether this week’s top Terminal-Bench model becomes available for self-hosting and what hardware it needs — the licence terms announced then will matter as much as the weights themselves. Second, independent evals of GPT-6.1 Terra outside OpenAI’s own benchmarks: MazeBench at 16% is the only non-vendor agentic data point available now, and closing that gap requires external testers.

Week’s verdict: Test GPT-6.1 Terra if you are currently paying Sol or Astra prices for agentic workloads and cost is a constraint. Run your own eval suite — not just Terminal-Bench — before committing.

Further Reading

• GPT-6.1 Terra benchmarks (benchlm.ai) — 18 benchmark rows with evidence ratings; six from independent leaderboards including Epoch AI FrontierMath v2.
• Kimi K3.1 Pro API and weights guide (aicybr.com) — pricing tiers, Zero Data Retention guidance, and what to verify before production use.
• How to Evaluate LLMs for Your Use Case (vortx.ch) — our pillar guide on building an eval suite before switching models, including how to move beyond a single leaderboard score.

Your turn: Which benchmark does your team actually use to make model-switching decisions — and does your production data back it up? Reply to our newsletter or send us a note — we feature the best answers in the Friday Scorecard.

How this article was made: AI researched and wrote this article under the Adrian persona, using the sources linked above, and it was published automatically without a human edit. Editorial guidelines are set by Adi. Spotted an error? Tell us and we will correct it.