Contents
The Week in One Paragraph
Three developments defined the week of October 24 to 30, 2026. OpenAI released GPT-6.1 Terra with API prices 50% below GPT-5.6 Terra and GPT-6.1 Sol and a large claimed gain on Terminal-Bench 4.0. Google announced Gemini 4 Carbon, but rollout is limited to cyber defenders and trusted partners, and developer access is still “coming soon.” Anthropic’s platform release notes raised the request size cap, launched a usage and cost analytics API, and set a November 30 removal date for fast mode on Claude Opus 4.8.
OpenAI GPT-6.1 Terra: a cheaper Terminal-Bench 4.0 contender
GPT-6.1 Terra (API id gpt-6.1-terra) is priced at $1 per million input tokens, $0.05 per million cached input, and $5 per million output, which OpenAI says is 50% below GPT-5.6 Terra and GPT-6.1 Sol. OpenAI reports 50.3% on Terminal-Bench 4.0 at maximum reasoning effort, against 21.5% for GPT-5.6 Terra. The release date is October 26, per the llm-stats tracker; OpenAI’s post says “starting today.”
What This Changes
- Cost per task. The headline is a per-token price cut, not a per-task one. OpenAI’s own Agents’ Last Exam comparison claims about 3x better token efficiency than Claude Opus 5.5 at roughly one-fourth the price per token, and a roughly 9x lower cost per task. Those components multiply to about 12x, not 9x, so read 9x as OpenAI’s headline figure rather than a derived one. Measure cost per completed task on your own workload.
- Reasoning effort is a cost lever. OpenAI’s internal error-inducing prompt set puts the error rate at 16.7% for Terra at low effort and 4.3% at maximum effort. Those are OpenAI’s numbers, but they show that effort settings change both cost and correctness, so an agent loop needs an explicit effort policy.
- Scope. OpenAI says GPT-6.1 Sol and GPT-6 Astra remain the better choice for the most complex coding tasks. Route by task difficulty, not by the benchmark that favors Terra.
- Access. Terra is in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and it is available through the API.
Google Gemini 4 Carbon: a frontier claim without developer access
Google announced Gemini 4 Carbon on October 28, 2026, and says it ranks #1 on Artificial Analysis Intelligence Index v5.0 with a score of 59. Google also reports SEC-bench Pro 88.2% for vulnerability finding, ExploitGym 54.8% for exploitation, and CWE-bench 71.0% for patching. Google describes Carbon as following Gemini 4 Argon, which it launched two weeks earlier.
What This Changes
- Deployment gating. This is the main change. Carbon starts with a subset of cyber defenders and expands through Google’s Fairwind Program. Google says it is “coming soon” to developers, businesses and consumers. Teams should not plan an API integration for Carbon this quarter on the strength of the announcement.
- Cost. Google has not published Carbon API pricing in the announcement, so no cost-per-task comparison is possible yet.
- Testing vendor claims. The index rank and the security benchmarks are Google-reported. The access boundary limits independent checking for now, so treat the scores as claims to verify once broader access opens.
- Use case fit. Security-adjacent teams, such as those doing vulnerability triage, are the group Google is prioritizing. Other teams should plan for evaluation readiness, not a model swap.
Anthropic’s platform changes: observability and a fast-mode deadline
Three first-party Claude API changes shipped in the same week. On October 28, the maximum request size for the Messages and Token Counting APIs rose from 32 MB to 128 MB. On October 29, a beta Usage and Cost Analytics API added endpoints at /v1/organizations/analytics/usage and /v1/organizations/analytics/cost, filterable by model, workspace, API key, and the user or service account behind each request. On October 30, fast mode for Claude Opus 4.8 was deprecated, with removal on November 30, 2026. After removal, requests to claude-opus-4-8 with speed "fast" return an error, and Anthropic’s migration target is Claude Opus 5.5.
What This Changes
- Cost attribution. The analytics API breaks spend down by the entity that caused it. That is the attribution layer needed to measure cost per task, and the older Usage and Cost API remains supported, so the change is additive for now.
- Hard deadline. The fast-mode removal is the one item here with a breaking change. Services that pin
speed: "fast"on Opus 4.8 need a regression test against Opus 5.5 before November 30. - Payload size. A 128 MB limit lets large repository or document payloads go in a single request. Budget those inputs by cost and latency, not only by whether they fit.
The Bottom Line
The week’s coding-AI competition moved from model quality toward price, access, and measurement. OpenAI is competing on per-token cost and cost per task while keeping its strongest model for hard work. Google is leading on claimed capability while holding developer access back. Anthropic is competing on operability: larger payloads, finer-grained cost data, and a forced migration off fast mode. For teams running coding agents, the actionable items are the November 30 deadline and a cost-per-completed-task measurement that you can run on your own workload.
Frequently Asked Questions
Can developers call Gemini 4 Carbon through an API today?
No. According to Google, Carbon is limited to a subset of cyber defenders, with phased expansion through the Fairwind Program, and developer availability is “coming soon.”
What does GPT-6.1 Terra cost per million tokens?
OpenAI lists $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens for gpt-6.1-terra. OpenAI says this is 50% lower than GPT-5.6 Terra and GPT-6.1 Sol.
What breaks on November 30 for Claude Opus 4.8 fast mode?
Per the Claude release notes, requests to claude-opus-4-8 with speed "fast" return an error after removal. Anthropic’s migration target is Claude Opus 5.5.
How This Briefing Was Researched
Sources were the primary vendor pages linked above: OpenAI’s GPT-6.1 Terra announcement, Google’s Gemini 4 Carbon announcement, and the Claude Platform release notes. The GPT-6.1 Terra release date comes from the llm-stats tracker, which cites the OpenAI post. Every vendor-reported benchmark and measurement is labeled as such. The authors did not run any of these models hands-on, and no independent benchmark was run for this briefing.
📖 Related Reads
- NoCode Insider — AI workflow automation with no-code tools, agents, and APIs
- ToolBrain — tool reviews, LLM comparisons, and AI workflow guides
Cross-links automatically generated from CodeIntel Log.
