Yes—Anthropic revoked OpenAI’s direct Claude API access in late July 2025, shortly before OpenAI launched GPT-5 on August 7. Anthropic said OpenAI technical staff had been using Claude Code and related tools and that the activity violated restrictions on building or training competing AI products. The public record does not prove that Claude outputs trained GPT-5, changed its capabilities, or that every route to Claude was blocked.
What happened, and when?
WIRED reported that Anthropic terminated OpenAI’s access on a Tuesday in late July 2025, identifying July 29 as the likely cutoff date. Anthropic’s spokesperson said OpenAI personnel had used Claude Code and other coding tools before GPT-5’s release and characterized that use as a direct violation of Anthropic’s terms. WIRED published its report on August 1; OpenAI announced GPT-5 on August 7.
As an Amazon Associate I earn from qualifying purchases.
| Date | Event | Evidence status |
|---|---|---|
| June–early July 2025 | OpenAI and Anthropic carried out a joint model-safety evaluation. | Confirmed by both organizations’ published reports. |
| Late July 2025 | Anthropic reportedly revoked OpenAI’s direct Claude API access. | WIRED report citing sources familiar with the decision. |
| August 1, 2025 | WIRED published the account of the cutoff and Anthropic’s explanation. | Published reporting. |
| August 7, 2025 | OpenAI publicly launched GPT-5. | Official OpenAI announcement. |
The reported action concerns Anthropic’s API access. It does not establish that Anthropic blocked every OpenAI employee from Claude.ai, Claude Code under every type of account, or Claude models offered through Amazon Bedrock, Google Cloud Vertex AI, or another reseller. Anthropic’s API documentation describes direct developer access, while its consumer-terms update distinguishes consumer services from commercial and API arrangements: API access documentation and consumer-terms update.
Free tools Windows power users keep installed
One-click scans. No signup required.
What Anthropic said OpenAI employees were doing
According to Anthropic’s account quoted by WIRED, OpenAI technical staff used Claude Code and related developer tools ahead of GPT-5’s launch. Reporting described activity involving coding, general task performance, safety behavior, refusal patterns, and comparisons with OpenAI systems.
Those details remain attributed allegations rather than independently documented test records. No public prompt logs, benchmark suite, user count, duration, or data-flow description shows precisely what OpenAI did. The public evidence also does not establish whether Claude responses became training examples, preference data, critiques, synthetic data, or merely internal observations.
That distinction matters. “Using Claude before GPT-5 launched” is supported by Anthropic’s statement as reported by WIRED. “Using Claude to test a finished GPT-5 model” is a narrower claim that has not been publicly demonstrated. “Using Claude to train GPT-5” is unproven.
What Anthropic’s terms prohibit
Anthropic’s commercial terms prohibit customers from using its services to build a competing product or service, including training competing AI models. They also prohibit reverse-engineering or duplicating Anthropic’s services. The relevant language appears in the Anthropic commercial terms, and WIRED described the same restrictions in its report.
Recommended Free Tools
Anthropic therefore said OpenAI’s particular use crossed a contractual line. The terms do not, on their face, prove that every comparison of Claude with another model is forbidden. A customer might benchmark models for procurement, compatibility, red-teaming, or safety research without using Claude to build a competing service. The public material does not disclose Anthropic’s internal test for separating permitted evaluation from prohibited competitive development.
Why the word “testing” is not enough
- Ordinary evaluation: measuring accuracy, latency, safety, refusal behavior, or reliability for a purchasing or research decision.
- Authorized safety testing: a documented exercise conducted under a joint agreement or written permission.
- Model-development use: feeding outputs into training, preference optimization, synthetic-data generation, prompt tuning, or system design for a competing model.
- Distillation or imitation: systematically extracting behavior or capabilities to reproduce another provider’s service.
The same API call can look like harmless benchmarking or prohibited development depending on authorization, scale, purpose, and what happens to the outputs afterward.
Does this prove Claude influenced GPT-5?
No. The chronology shows that the reported cutoff preceded GPT-5’s public release, and Anthropic explicitly connected the alleged use to the prelaunch period. It does not show that Claude outputs were included in GPT-5’s training data, that Anthropic discovered a GPT-5-specific extraction program, or that the cutoff altered GPT-5’s final performance.
OpenAI’s GPT-5 announcement and system card describe a unified system combining fast responses, deeper reasoning, and routing across model variants. They do not identify Claude as a training or evaluation dependency. The defensible formulation is: Anthropic said OpenAI employees were using Claude before GPT-5’s launch; no public evidence establishes that Claude outputs trained GPT-5.
How the cutoff fits with the companies’ joint safety work
The companies had recently cooperated. In June and early July 2025, they conducted a limited alignment-evaluation exercise in which each organization tested the other’s publicly released models. Anthropic’s report says OpenAI models were accessed through the Anthropic API, and GPT-5 was not available during that period. Anthropic published its account at alignment.anthropic.com; OpenAI published its account at openai.com.
That is not necessarily inconsistent with the later termination. A mutually agreed safety exercise can authorize activities that ordinary commercial terms would not. Neither organization’s public report, however, spells out whether the joint agreement created permissions different from standard customer contracts.
The episode thus reflects a tension rather than a complete collapse of cooperation: frontier labs may need cross-model testing to find safety failures, while also trying to prevent competitors from using their models for imitation, distillation, or product development.
What remains unresolved
- OpenAI has not, in the sources available here, provided a detailed public rebuttal to Anthropic’s specific explanation.
- No public record identifies the OpenAI account, team, prompts, benchmark files, volume, or duration involved.
- There is no public proof that Claude outputs entered GPT-5 training or changed GPT-5’s quality.
- The reported action does not establish that consumer Claude accounts or all third-party hosting channels were blocked.
- The public record does not state whether the suspension was permanent, how broad it was, or whether access could be restored under different terms.
Why this matters for developers and enterprise buyers
Terms can govern competitive benchmarking
Organizations should not assume that calling an activity a “benchmark” makes it permissible. Before running automated or high-volume comparisons, obtain written confirmation of the provider’s rules and document the intended use.
Keep an auditable data trail
Record the purpose of each evaluation, authorization, prompts, output retention, and whether responses enter training, fine-tuning, preference, or synthetic-data pipelines. This is especially important when one model provider is a direct competitor of another.
Cloud marketplaces do not automatically remove restrictions
Amazon Bedrock and Google Cloud Vertex AI can centralize identity, billing, and governance, but hosting through a reseller does not automatically override the underlying model provider’s acceptable-use or commercial terms. Check the contract that applies to the specific region, model, and account.
Plan for access changes
A provider can rate-limit or terminate access after an abuse review. Anthropic had also announced Claude Code rate-limit changes amid explosive usage and terms concerns, although the public record does not prove that those changes and OpenAI’s termination were one coordinated action. Critical evaluations should have an approved fallback rather than depend on a competitor’s API.
The bottom line on the GPT-5 headline
Anthropic did revoke OpenAI’s direct Claude API access before GPT-5 launched, and Anthropic said OpenAI staff had used Claude tools in violation of its commercial terms. What the incident does not prove is that Claude trained GPT-5, that Anthropic blocked every form of Claude access, or that the cutoff changed GPT-5. The lasting issue is how frontier-model companies draw a contractual boundary between legitimate cross-model safety work and competitive model development.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
