AI
“OpenAI and Anthropic reportedly came close to signing a legally binding agreement to allow API access for testing each other's public AI models for safety risks, though it remains unclear if the deal was finalized" (post image text: "OpenAI and Anthropic reportedly working on a deal to stress-test each other's AI models")”
Plain restatementThe Information reported that OpenAI and Anthropic negotiated, and came close to concluding, a legally binding agreement under which each would receive API access to the other's commercially available models in order to run its own safety and vulnerability tests, with no retention of the other party's data, and it is not known whether the agreement was signed.
This post accurately summarizes a real news report. The Information reported on September 21, 2026 that OpenAI and Anthropic negotiated, and came close to concluding, a legally binding agreement giving each company API access to test the other's publicly available models for safety problems, with neither side allowed to keep the other's data. The post is right that it is unknown whether the agreement was ever signed, and it correctly credits the source. What the post does not say is that the entire story rests on one unnamed source at one outlet, that every other article you may see about it is just repeating that same report, and that neither OpenAI nor Anthropic has commented. The image headline saying the companies are "working on a deal" is also looser than the reporting, which describes talks that happened earlier in the year with an unknown outcome. The background detail is solid: OpenAI and Anthropic really did run a joint safety evaluation of each other's models and published the results in August 2025. Treat the deal itself as a credible but unconfirmed report, not as a thing that happened.
OpenAI and Anthropic [drifted from the evidence:] reportedly came close to [drifted from the evidence:] signing a legally binding agreement [drifted from the evidence:] to allow API access [drifted from the evidence:] for testing each other's [drifted from the evidence:] public AI models [drifted from the evidence:] for safety [drifted from the evidence:] risks, though it [drifted from the evidence:] remains unclear if the [drifted from the evidence:] deal was [drifted from the evidence:] finalized" (post image text: "OpenAI and Anthropic reportedly working on a deal to stress-test each other's AI models")
[added by the neutral restatement:] The Information reported that OpenAI and Anthropic [added by the neutral restatement:] negotiated, and came close to [added by the neutral restatement:] concluding, a legally binding agreement [added by the neutral restatement:] under which each would receive API access [added by the neutral restatement:] to the other's [added by the neutral restatement:] commercially available models [added by the neutral restatement:] in order to run its own safety [added by the neutral restatement:] and vulnerability tests, with no retention of the other party's data, and it [added by the neutral restatement:] is not known whether the [added by the neutral restatement:] agreement was [added by the neutral restatement:] signed.
Red-tinted words in the claim drifted from the evidence. Green-tinted words are what a neutral restatement needs.
The trace / claim to source
- The Information did publish this report, on 2026-09-21, as an exclusive by Amir Efrati and Stephanie Palazzolo. The post's attribution to The Information is correct.
- The described terms match the reporting: mutual API access, commercially available models only, and a no-data-retention commitment.
- The post's statement that it is unclear whether the deal was completed matches the source exactly. This is the single most important thing the post got right, and most viral restatements of deal rumors omit it.
- The word "reportedly" is used, correctly signaling that this is a report rather than a confirmed event.
- The background sentence is substantively correct. A joint OpenAI-Anthropic safety evaluation did occur and was published on 2025-08-27 by both companies, and it surfaced different weaknesses in each side's models, including deception-related and misuse-related behaviors.
- Date context mismatch: the on-image headline says the companies are "working on a deal," present and ongoing, while the reporting describes negotiations that took place earlier in 2026 with an unknown outcome and no indication of current activity. The caption's "came close to signing" is accurate; the headline card is not, and headline cards are what most viewers read.
- Omitted qualifier: the post does not disclose that the entire report rests on one unnamed source at one outlet, and that neither OpenAI nor Anthropic has commented. A reader has no way to gauge how thin the sourcing is.
- Omitted qualifier (secondary, background sentence): the 2025 joint evaluation results were published by both labs with explicit cautions that they came from limited synthetic scenarios with some safeguards deliberately relaxed, and were not claims about real-world product behavior. "Found different problems in their models, including misleading behavior and harmful responses" is a defensible one-line compression but drops those conditions.
- Whether the agreement was ever signed. This is unknown to the reporter as well, not just to me.
- Whether negotiations are still live as of 2026-09-24. No source states the current status.
- The full text of The Information's article is paywalled. I retrieved the headline, the outlet's own social summary, and multiple independent syndications that agree on the details, but I did not read the original in full, so exact wording of the terms and of the sourcing description comes from summaries rather than the original.
- Whether the negotiated agreement, if signed, would have covered the models involved in the 2026 security incidents that coverage places nearby. Coverage explicitly says the timing relationship is unclear.
- Whether either company will confirm, deny, or announce anything later.
The Information published an exclusive on 2026-09-21 reporting that OpenAI and Anthropic, working with their respective lawyers earlier in 2026, neared a legally binding agreement to stress-test each other's AI models for vulnerabilities and hidden dangers. Secondary summaries of that article consistently describe the proposed terms as: mutual API access to each company's commercially available models, exclusion of unreleased models, and a commitment by each side not to retain the other's data. Every summary states it is unclear whether the agreement was finalized. The reporting is attributed to a single unnamed source described as a person with direct knowledge of the talks. Coverage notes the talks predated a series of cybersecurity incidents involving OpenAI's technology disclosed in 2026, and at least one summary states neither company would comment, with another outlet reporting that neither responded to a comment request. The background element is separately verifiable from primary sources. On 2025-08-27 Anthropic and OpenAI each published findings from a pilot cross-evaluation in which each lab ran its own alignment and misuse evaluations on the other's publicly available models over the other's public API, with certain external safeguards relaxed. OpenAI evaluated Claude Opus 4 and Claude Sonnet 4; Anthropic evaluated GPT-4o, GPT-4.1, o3 and o4-mini. Reported topic areas included sycophancy, deception and scheming, self-preservation, hallucination, instruction hierarchy, and jailbreak resistance. Both labs characterized the results as early-stage findings from limited synthetic scenarios and cautioned against broad cross-provider conclusions.
Complete reasoning
The reply is formatted for pasting into the thread where the claim is circulating.
Compact share page: ai.trueseeker.com/s/09a3d2ea8dc5/CkMDOGqGBf8kRX9uBGk4fRN58H5
Ask this case
Answers come only from the case file above; nothing is added.
Did OpenAI and Anthropic actually sign a deal to test each other's AI models?
It's unclear. The Information reported the two companies came close to a legally binding agreement, but no source confirms whether it was ever finalized or signed.
What would the agreement have allowed, according to the report?
Each company would get API access to the other's commercially available (not unreleased) models to run their own safety tests, with a commitment that neither side would keep the other's data.
How solid is this report? Who is the source?
It comes from a single unnamed source described as having direct knowledge of the talks, published as an exclusive by The Information. Neither OpenAI nor Anthropic has commented on it.
Is the claim that they are currently 'working on a deal' accurate?
That framing is looser than the actual reporting. The negotiations described took place earlier in 2026, and no source confirms the talks are still ongoing.
Have OpenAI and Anthropic done anything like this before?
Yes. In August 2025 both companies published results from a joint pilot evaluation where each tested the other's public models for issues like deception and jailbreak resistance, though both cautioned the findings came from limited, synthetic test scenarios.