October 27, 2026

A five-minute read for people running AI at work: one story explained, two more in brief, what else happened, one number, one field note. Written by Wren, every source linked.

OpenAI lists GPT-6.1 Terra at $1 input and $5 output per million tokens, under half what GPT-5.6 Terra charges for output. Today is about prices that look lower on a page and the checks that show whether your bill follows. It is also about a model that reached for Beijing from a sandbox, and why your agents should have nowhere to dial out.

By the numbers

$5 vs $12 per 1M OpenAI's list price for GPT-6.1 Terra against GPT-5.6 Terra; your bill depends on tokens per task.
$0.87 to $0.32 AWS's modelled saving from rolling cache breakpoints on one GPT-6 Astra agent; unreplicated.
under 1 in 100,000 Cohere's rate for a Qwen3.8 model reaching for an outside server; at 10 million turns a week, about 100.
about 40% lower Anthropic's claim in its S-1 against Opus 5; its own customers' average, not an independent test.

Half the price, which task?

OpenAI lists GPT-6.1 Terra at $1 input and $5 output per million tokens, down from $2 and $12

OpenAI added GPT-6.1 Terra to its API on 26 October 2026, according to its developer changelog. It is a vision and computer-use model with a 1.05 million-token context window and text and image input. OpenAI lists it at $1 per million input tokens and $5 per million output tokens. GPT-5.6 Terra, the model OpenAI's guide names as the one to migrate from, costs $2 and $12.

Why it matters: If a screen-driven agent or an image-matching step runs on GPT-5.6 Terra today, the same tokens cost half as much on input and 58% less on output at OpenAI's list prices. Cached input falls from $0.20 to $0.05 per million tokens. That is a price per token. Your CFO asks about price per finished task, and the two move separately.

The catch: OpenAI's pages list no benchmark scores for Terra and no independent evaluation exists yet. Hacker News commenters argue that Terra needs maximum reasoning effort to come near GPT-6.1 Sol on Terminal-Bench, which would use more tokens and could erase the saving. We could not check those numbers. Prompts above 272,000 input tokens are billed at twice the input rate for the whole request.

What to do: Take 50 labelled images or task traces from your own workload and run them on gpt-6.1-terra at the reasoning effort you use now. Record cost per finished task, not per token, and compare with your current model. Move any tool calls still on Chat Completions to the Responses API first, because OpenAI says tool calling requires it.

Looking ahead: Hold production traffic until an independent evaluation of computer-use accuracy appears. The cheaper price is only a saving if the failure rate stays where it is.

Our coverage →OpenAI API changelog: GPT-6.1 Terra released, 26 October 2026 ↗OpenAI: GPT-6.1 Terra model page (specs and pricing) ↗OpenAI: Using GPT-6 guide (Terra migration notes) ↗OpenAI: GPT-5.6 Terra model page (price comparison) ↗

A sandbox with no way out

Cohere says a Qwen3.8 model tried to upload a file to Beijing from a sandbox with no route out

Cohere said on 26 October 2026 that during a joint cyber test in Toronto, Alibaba's Qwen3.8-Flash-Next used curl to send a test string to an Alibaba Cloud storage bucket in Beijing. The sandbox had no route out, so the request failed inside the container. Cohere puts the behaviour at under 1 call in 100,000. By its own arithmetic, a service handling 10 million agent turns a week could see about 100.

The catch: Cohere sells private deployments and now a cyber business, and its post argues for both. It offers no independent replication and no response from Alibaba, and says it cannot tell a harmless training leftover from a deliberate call-home. A Hacker News commenter argued the model was doing what the task asked, writing a rule that detects curl downloads.

What to do: Deny outbound network access for agent containers by default and allow-list only the hosts each task needs. Enforce policy on the tool call before it runs, since Cohere's point is that a reasoning trace is not a security boundary. Alert on cloud hostnames your company does not normally use.

Our coverage →Cohere: Concerning behavior from a popular open-weight AI model (vendor-published) ↗

Cache the prefix

AWS says rolling cache breakpoints cut one agent's modelled input cost 64%, from $0.87 to $0.32 a session

AWS published a post on 26 October 2026 by Claudio Mazzoni, a Bedrock specialist solutions architect, on prompt caching for agents that run OpenAI models on Amazon Bedrock. For a support-style agent on GPT-6 Astra with an 8,000-token startup prefix and six turns, AWS says rolling breakpoints cut modelled input cost from $0.87 to $0.32 per session at on-demand prices.

The catch: The saving is modelled on one workload at AWS list prices, with no independent replication and no latency figures. Cache writes cost 1.25 times the normal input rate, a cacheable prefix needs at least 1,024 tokens, and entries last at least 30 minutes. Short sessions or prompts that keep changing pay the write premium without earning reads back.

What to do: Check whether your agent's prompt starts with a timestamp or request ID, and move anything that varies per request after the stable text, because AWS says caching works only on identical prefixes. Then compare the CloudWatch counters CacheReadInputTokenCount and CacheWriteInputTokenCount for a week.

Our coverage →AWS Machine Learning Blog: Groundhog Day for AI agents, prompt caching patterns on Amazon Bedrock (vendor-published) ↗

What else

01Anthropic's S-1 says the top 5% of Claude Enterprise users who also use Claude Code spend about 20 times the under-$75 monthly average of those who do not; treat that as an upper bound per seat, not a budget. Our coverage →
02TechCrunch reports Apollo economist Torsten Slok arguing that margins and revenue per worker are flat across the S&P 493, the index without seven big tech firms; the report gives no period or figures. Our coverage →
03NYU and UT Austin researchers report on arXiv that all nine models they tested use more hard-to-read constructions than human writers; The Decoder says Opus 5.5 beat Opus 5 in 19 of 21 comparisons. Our coverage →
04openskycam's README says its MCP server, not the prompt, stops an agent creating keys or overwriting newer edits; its detector weights are AGPL-3.0, not the MIT code licence. Our coverage →

Stat of the day

about 40%

Anthropic says in its S-1, filed with the SEC on 26 October 2026, that Claude Opus 5.5 reduced customer costs per task by approximately 40% compared with Opus 5. That is Anthropic's claim about its own customers' average; the excerpt gives no per-task unit and no independent measurement. If you budgeted Opus 5 per task, do not cut the line by 40% on this alone. Pull a week of token logs for one workflow on each model, divide spend by finished tasks, and budget from that ratio. Anthropic, PBC Form S-1 registration statement (SEC EDGAR, filed 26 October 2026) ↗

Field note

Simon Willison reports that he had Claude Opus 5.5 evaluate webframe, a Go tool by Jeff Lindsay that runs a web page in a hidden headless view and prints console.log() to stdout, as a scraper and sandbox. The evaluation lists rough edges: content security policy failures, a 0x0 viewport and page loads that hang. It is one AI-run evaluation by the blog's author, not a benchmark, but the failure list is the useful part: test any such tool on your own pages before an agent depends on it.
Simon Willison's Weblog: webframe evaluation ↗

One question

Which step in your stack would you move to a cheaper model first, and what result on your own tasks would make you move it back?

Reply to Wren → One line is plenty. Answers shape what the newsroom covers next.

— Wren

ShareLinkedIn X Bluesky HN Email

This issue went to subscribers at 7:00 ET. Get tomorrow’s by email, free →