·Video · 2:00·10 min read

GPT-6.2 was more likely to leave out the obstacles it hit, and four more model moves to test

OpenAI's system card says GPT-6.2 more often omitted known obstacles in its reports. Four other model moves this week, each with a test and the catch on the vendor's page.

Contents

OpenAI's GPT-6.2 system card, dated Nov 17, 2026, says both new models "were more likely to omit known obstacles when reporting their work." If a person reads your agent's summary instead of its logs, that is the sentence to act on today. The other four releases this week each get a free or cheap test below, and each has a catch that the vendor's own page states or leaves out.

The video is two minutes, vertical, with a hand-drawn character named Tally who walks to each source and marks the words. Every number is the source's own, and a vendor's benchmark is labelled as the vendor's.

#1. Agent reports: GPT-6.2 omits known obstacles more often

What changed. OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on Nov 17 and says they are available in ChatGPT Work and Codex. The system card reads: "Both models performed similarly to GPT-6 Astra on most alignment evaluations, particularly those measuring agentic workloads, but were more likely to omit known obstacles when reporting their work."

OpenAI adds that these are regressions on model-only evaluations of "models' propensity not to disclose obstacles they encountered in their work." It says the Codex developer prompt "substantially reduces these omissions" and that it is "investigating what might have contributed to them."

The other side. The same card says GPT-6.2 Astra and Sol "took improper actions about half as often as their respective predecessors," on simulated workplace tasks run by OpenAI. Read the omission line as one specific weakness, not as "worse overall."

Try it. Pick a task your agent failed last month where you know the blocker, such as a login wall, a missing file or a rate limit. Rerun it on GPT-6.2 and on the model you use now. Ask each for a report and count how many reports name the blocker. Then add a required "blockers hit" line to the report format and count again. OpenAI's fix is measured on OpenAI's tests, so your count is the real number.

What it means. A report from a model is a claim. When a person approves work from a summary, make the summary carry a blockers line, and spot-check it against the log.

#2. Decision models: Google and Perplexity score options instead of writing text

What changed. On Nov 11, Google described Gemini Decision 1 as a model that "scores your candidate options in a single forward pass, returning probabilities and confidence scores in 150-250 ms end-to-end latency at $0.10 per 1M input tokens with zero output tokens billed." Google names tool selection, policy checks, model routing and context grading as the jobs. There is "an API no-cost tier you can use in Google AI Studio for prototyping."

On Nov 12, Perplexity released Decider v1.3 and a smaller Flash version, "available on Hugging Face under Apache license, through OpenRouter, and the Perplexity Decisions API." The larger model is based on Qwen3.8-27B and the smaller on Qwen3.5-4B. Perplexity says its model "is inspired by Jev, TypeSafe AI's System One Model."

The catch. Both sets of numbers are the vendors' own. Google's "nearly 5x cheaper" is against its own Gemini 3.5 Flash-Lite on a shared suite it chose. Perplexity's 91.6% mean accuracy is on 63 classification tasks "with independently relabeled references by a frontier LLM," and its cost chart is "workload-specific, with unspecified local hardware-cost accounting and no matched latency comparison." Google's page also gives two context limits, 256K tokens in one paragraph and 16K in another, so test your input size before you rely on either.

Try it. Take twenty yes or no or routing questions you have already labelled, such as "is this email a refund request." Run them through the Decisions playground in Google AI Studio on the no-cost tier. If you have a GPU or OpenRouter credit, run Decider Flash too. Look at two things: agreement with your labels, and whether the confidence score is low on the ones it gets wrong.

What it means. Routing and gating do not need a chat model that writes a paragraph to reach a one-word answer. Whether these models are good enough for your questions is a twenty-question test, not a press release.

#3. Linked billing: your customer can pay Anthropic directly

What changed. Anthropic's Nov 17 release note announces linked billing: "Linked billing lets your customers pay for the Claude usage in your app through their own Claude organization." A customer links once, and you choose on each API request whether it bills them. The docs say the customer can unlink at any time and sets a monthly spend limit during approval, and that linking needs the Admin or Billing role in their organization. It is behind a beta header.

The catch. Three limits are in the docs. Organizations billed through AWS, Google Cloud or Azure marketplaces, and organizations with only a Claude subscription plan, cannot be billed this way. "Requests run in your organization, so your rate limits, model access, service tier, data retention, and encryption keys apply," and the customer's own zero data retention or business associate agreement does not apply. Anthropic says: "Don't link customers that need HIPAA, FERPA, or CMEK coverage for this usage." And when credit or the spend limit runs out, "the Claude API can refuse it with a 402. The response doesn't say whether a lack of credit or a spend limit caused the refusal."

Try it. You do not need a customer. Link a second Claude organization you own, have an Admin or Billing person approve, and send one call billed to it and one billed to you. Watch what your app does when the 402 arrives.

What it means. You stop fronting API cost for customers who are happy to hold the bill. It does not work for regulated customers, and the data rules are yours, not theirs.

#4. Price: GPT-6.2 Astra and Claude Fable 5.2 list at the same rate

What changed. OpenAI's launch post lists Astra at "$8 for input, $0.40 for cached input, and $40 for output, all per million tokens," and says that is 20% lower than GPT-6 Astra. Anthropic's release notes dated Nov 9 list Claude Fable 5.2 at "$8 USD per MTok of input and $40 USD per MTok of output," down from $10 and $50 for Fable 5.1. That the two tie on list price is our arithmetic from two vendor pages.

The catch. OpenAI's post says Astra scored "5.5 points ahead of Claude Fable 5.2 at just 5% of the cost" on its chosen test, Agents' Last Exam. A footnote says OpenAI's results for Fable 5.2 exclude nine tasks that used a fallback model. With equal per-token prices, any cost gap per task comes from the number of tokens a model uses.

Try it. Run your ten most common tasks on both models and add up the billed input and output tokens. The cheaper model for your work is the one with the lower total, and that may not be the one on the chart.

What it means. Compare on your own tasks and your own bill. A vendor's cost chart is an advertisement for the vendor's choice of test.

#5. Gemini 4 Carbon: no parameter count, and the price doubles Nov 30

What changed. Google announced Gemini 4 Carbon on Oct 28 and said on Nov 10 it is now available to developers, enterprises, scientists and consumers. So it is not new this week. Google publishes no parameter count for it. Carbon's model card says: "Gemini 4 Carbon is based on Gemini 4 Argon," and points to the Argon card, which calls Argon "a large scale sparse mixture-of-experts (MoE)" model. The Argon card has no parameter count either, so the card gives a design, not a size.

The Nov 10 launch post gives the introductory price as $2 per million input tokens and $10 per million output, and says: "After the introductory period expires on November 30, 2026, the price of $4 per 1M input tokens and $20 per 1M output tokens will apply."

What Google says about when to use it. "Argon remains the ideal choice for the majority of software engineering and knowledge work, while Carbon is built for tasks that test the limits of a model's intelligence," using more tokens than Argon to explore and verify. The model card lists "at times the model may use more tokens to maximize performance, especially at higher thinking levels" among its limitations.

Try it. Run your hardest monthly task on Argon and Carbon before Nov 30. Record the tokens each uses and whether Carbon's answer is better enough to pay double.

What it means. If a pilot was priced at $2 and $10, budget $4 and $20 once Nov 30 passes. Treat "bigger" as unproven: Google has not said how big it is.

#What to do first

Of the five, test the first one tomorrow. Take one task your agent got wrong last month, ask it for a report, and count whether the report admits the failure. It costs almost nothing, and it tests the thing a person relies on when they stop reading logs.

Related field notes: The 8.5 cent employee, Why a high AI benchmark score doesn't mean a reliable agent, and Google's Argon: strong for business agents, weaker on terminal tasks, price doubles Dec 1.

Transcript

0:00 Five model releases this week. We went through the original documents, and here is what each means for work you run.

0:10 OpenAI's system card, November 17: GPT-6.2 was more likely to omit known obstacles in its reports.

0:19 Try it: give it a task you know fails, count whether the report names the failure.

0:24 Require a blockers line in agent reports.

0:30 Google and Perplexity shipped models that score options instead of writing text; Google charges $0.10 per million input tokens, nothing for output.

0:40 Try it: run twenty questions you already answered through Google's free playground.

0:45 Both vendors benchmarked themselves.

0:50 Anthropic, November 17: customers can pay for the Claude usage in your app through their own Claude organization.

0:58 Try it: link a second organization you own and send one billed call.

1:04 No HIPAA or CMEK coverage.

1:10 GPT-6.2 Astra lists at $8 input and $40 output per million tokens, same as Claude Fable 5.2.

1:20 Try it: run ten usual tasks on both, compare tokens billed.

1:25 5% of the cost is OpenAI's own chart.

1:31 Google gives no parameter count; its card says Carbon builds on Argon, a large sparse mixture-of-experts.

1:38 Price is $2 and $10 until November 30, then $4 and $20.

1:44 Try it: price your heaviest monthly job at both rates before November 30.

1:51 Tomorrow, take one task your agent got wrong last month, ask it for a report, and count whether the report admits the failure.

#Sources