AI Latest News: 29 October 2026, API Costs and Access

GPT-6.1 Terra’s lower API prices put workload cost and access at the center of AI news for the week of 29 October 2026. The announcements from 16-29 October offer different paths to production: cheap APIs, downloadable models and restricted previews.

Lower API prices make workload testing more valuable than model rankings

OpenAI introduced GPT-6.1 Terra for visual reasoning and computer use, with lower API prices than its predecessor. MiniMax’s 19 October M3.1 Flash announcement adds another inexpensive option, while Google has published introductory pricing for its restricted Carbon release.

Test MiniMax first for narrowly defined, cost-sensitive text workloads, and shortlist Terra when visual reasoning or computer use is central. These announcements don't establish equivalent accuracy, latency or tool reliability. Compare candidates against the same inputs and acceptance criteria, using cost per accepted result: another model call or manual correction can erase the token-price advantage.

Kimi’s API release and promised weights are separate milestones

Moonshot’s Kimi K3.1 Pro announcement makes Pro and Highspeed available through Kimi products and the API. Full model weights are promised for a later date.

The published details are:

• Context window: one million tokens.
• Current access: Kimi products and API.
• Promised full-weights release: 4 November.

A large context window makes Kimi worth testing on document-heavy tasks, but capacity alone doesn’t establish retrieval accuracy or consistent reasoning across the input. A contract-review test should check whether the model finds conflicting clauses across distant documents.

Keep self-hosting commitments pending until the weights arrive and your team has inspected the license, hardware requirements and serving instructions.

DeepSeek’s releases improve control, but leave production work to you

DeepSeek’s 17 October release updates its coding model to V4.1-Flash-1017 and opens Search and Fetch APIs. Separately, the DeepInfer repository exposes an inference framework used behind DeepSeek’s production services, with components deliberately excluded.

Implementation control is the strongest reason to inspect DeepInfer; cheaper self-hosting isn't a given. Its README excludes the KV-cache system and load balancing needed for a complete inference service. Those omissions matter when requests compete for memory and compute.

JSON-schema-constrained output helps a batch extraction job produce records in a required structure, but it doesn’t establish that the extracted values are correct. Start with an offline workload whose outputs can be checked before considering an interactive service.

For preserving progress across interrupted agent runs, see Pi Durable in Practice: Building Agents That Resume, which covers durable workflow state separately from inference execution.

Liquid AI makes local deployment more plausible by narrowing the job

Liquid AI’s 27 October LFM3 release introduces six specialized models derived from a shared backbone, with support for single-GPU fine-tuning. The LFM3-1.7B model card states both deployment characteristics and discouraged tasks.

Its published specifications and boundaries include:

• Context window: 256K tokens.
• Approximate model size at 4-bit precision: 1 GB.
• License: LFM Open License v1.0.
• Discouraged uses: agentic coding and complex multi-step reasoning.

A small model is worth testing for a local document classifier or narrowly scoped assistant. No benchmark result is reported for those candidate workloads. The approximate weight footprint also leaves application memory unaccounted for, and the license needs review for your intended distribution or commercial use.

Gemini 4 Carbon is a restricted preview, not a procurement option

Google’s 28 October Gemini 4 Carbon announcement starts access with selected cyber defenders through the Fairwind Program. Broader developer access is still forthcoming; introductory pricing doesn't mean your team can deploy it now.

Put Carbon in a restricted-preview category in internal comparisons, with no assumed migration date. Security teams need to establish eligibility and access before evaluating it as a production replacement.

Worth watching

• Kimi weights: Check whether the promised release appears by 4 November, then inspect its license and deployment documentation in the official announcement.
• DeepSeek tool billing: Confirm current Search and Fetch charging terms in the release notice before expanding tool-heavy workloads.
• Carbon availability: Look for an explicit broader-access announcement from Google.

Your Next Move

1. Run the same 30 representative tasks through Terra and MiniMax. Record accepted outputs, total token charges and manual corrections.
2. Check your DeepSeek integration against the recommended Harness release, then count Search and Fetch calls separately from model tokens.
3. Before testing DeepInfer, check whether your target hardware matches its documented support and assign ownership of the missing serving components.
4. Test LFM3-1.7B on one bounded local task, measuring peak application memory and task accuracy. Exclude agentic coding.
5. Put Kimi’s promised weights release on your 4 November checklist; keep self-hosting commitments pending until the artifacts and terms can be inspected.

Useful Resources

• GPT-6.1 Terra announcement - API identifier, capabilities and published pricing.
• DeepInfer repository - Supported hardware, offline inference scope and missing service components.
• LFM3-1.7B model card - Context capacity, license and discouraged use cases.
• Kimi K3.1 Pro announcement - Current API access and promised weights-release date.

Looking Ahead

Choose an execution path your team can access, afford and operate for a specific job. An inaccessible model shouldn't delay a decision your team can test against available services today.