Concede the basic point before the analysis: the OpenAI and Anthropic renegotiation of their pre-deployment evaluation agreements with CAISI is a private process. The specific terms under negotiation are not public, the timeline is not public, and any account of what is happening relies on partial signals — public statements from the labs, public statements from the administration, and the May 5 2026 agreements with Google, Microsoft, and xAI as the comparison template.
That said, the renegotiation matters and the structural shape is visible. OpenAI and Anthropic had pre-deployment evaluation agreements with the original AISI signed in 2024 under the Biden-era Executive Order 14110. The executive order has been substantially revised. AISI has been restructured into CAISI. The political posture of the federal AI evaluation regime has shifted toward a more cooperative-with-industry stance. The two labs that are missing from the May 5 signing are the two labs with the strongest historical ties to the prior administration's approach, and the renegotiation is the explicit mechanism for resolving the tension.
Three possible outcomes. The shape of each outcome would affect frontier model release governance for the next 24 months.
Outcome 1: Agreements Renew Under May 5-Style Terms
The most likely outcome based on publicly visible signals. OpenAI and Anthropic sign agreements substantially similar to the May 5 template: pre-deployment access window, narrowed scope to national-security-relevant capabilities, private-first findings, voluntary structure with no enforcement mechanism.
This outcome resolves the immediate uncertainty without significant policy change. Both labs continue to ship frontier models on roughly current cadences. The federal evaluation regime achieves coverage across all major US-headquartered frontier labs. The political posture is settled.
The strategic logic for the labs to sign under May 5-style terms: the alternative (walking away from federal evaluation entirely) carries political and reputational risk. Both labs have publicly committed to safety as a value, and walking away from voluntary federal evaluation would create dissonance with that commitment. The May 5 terms are also low-friction enough that signing does not impose significant operational cost.
The strategic logic for the administration to accept these signings: getting both labs into the regime expands the federal evaluation footprint and demonstrates that the more cooperative posture is workable. Refusing to renew under the May 5 terms would create an adversarial dynamic that does not serve administration goals.
If this outcome lands, the announcement is likely inside the next two release cycles — somewhere between June and September 2026. The next major OpenAI or Anthropic release (probably the next Opus release in Q3 or a GPT-5.6/6.0 release in Q3-Q4) becomes the natural milestone for the agreement to be in place.
Outcome 2: Agreements Lapse, Labs Continue Without Federal Evaluation
A possible but less likely outcome. Negotiations fail, OpenAI and Anthropic continue to develop and release models without federal pre-deployment evaluation. Both labs maintain their internal safety evaluation processes, work with the UK AISI, and possibly engage with the EU AI Office under the EU AI Act framework.
This outcome would be politically costly for both labs and would create a multi-tier US frontier model market — Google, Microsoft, and xAI under federal evaluation, OpenAI and Anthropic operating outside it. The asymmetry might create competitive consequences for the latter two labs (e.g., federal contracting preferences, public perception around safety) that the labs would absorb in exchange for greater operational freedom.
The strategic logic for this outcome to be pursued: only meaningful if OpenAI or Anthropic determine that the May 5 terms compromise something they value highly — most likely the scope of evaluation (if the administration pushes for narrower scope than the labs prefer) or the disclosure structure (if the administration pushes for more or less public reporting than the labs prefer). Without a substantive disagreement on terms, walking away does not serve either lab's interests.
The probability of this outcome is low based on visible signals. Both labs have indicated public willingness to engage. The administration has not signalled hardball negotiation tactics. But the outcome is non-zero and would have meaningful consequences for the broader AI governance picture.
Outcome 3: Agreements Renew With Custom Terms
The intermediate outcome. OpenAI and Anthropic sign agreements that differ from the May 5 template in specific ways that reflect their priorities and concerns. Possible custom terms:
- Different scope: Either narrower (focused only on CBRN uplift) or broader (including some bias and misuse concerns) than the May 5 template. - Different disclosure structure: Either more aggressive public reporting (closer to the original AISI structure) or less aggressive (private-only with no aggregate public reporting). - Different evaluation window: Either longer than 30 days (more comprehensive evaluation) or shorter (faster release cycles). - Reciprocal commitments: The agreements include federal-side commitments around export controls, federal procurement, or AI Action Plan implementation that benefit the signing labs.
Custom terms would signal that the federal evaluation regime is moving toward differentiated agreements rather than a uniform template. This has both benefits (each lab gets terms that fit its specific situation) and risks (the regime becomes harder to manage, less predictable for other labs that follow).
Probability of this outcome: moderate. The labs have historically negotiated for terms that fit their specific concerns, and there is no obvious reason they would accept a one-size-fits-all template if better terms are achievable.
What The Renegotiation Means For Builders
Two practical implications for developers and teams building on top of OpenAI and Anthropic models.
Release cadence uncertainty: Until the renegotiation resolves, the timing of major OpenAI and Anthropic releases carries a small political-overhead risk. The next Opus release (expected Q3 2026) could be delayed by weeks if the renegotiation is unresolved at that point. Similarly for the next major OpenAI release. The likely effect is small but worth tracking for teams with release-dependent product roadmaps.
Policy-driven feature changes: The renegotiation could include terms about specific model behaviours — refusal patterns, capability gating, safety filtering thresholds. If the agreements impose specific behaviour requirements, the underlying models may ship with different default behaviours than they would have absent the agreements. This is unlikely but not impossible.
For most builders, the renegotiation is a background condition that does not require active management. The base case is that agreements renew under May 5-style terms and the underlying model availability continues unchanged. Tracking the announcements is sufficient.
The Anthropic Specific Angle
Anthropic's position in this renegotiation is structurally different from OpenAI's. The lab has been publicly more vocal about safety evaluation, has historically advocated for stronger federal oversight, and has incorporated evaluation findings more deeply into its release decisions than other labs. The Anthropic posture suggests the lab will sign a renewal under most plausible terms.
The interesting question for Anthropic is whether the renewal includes the disclosure provisions that the lab has historically favoured. The original AISI structure allowed for more public reporting of findings than the May 5 template appears to permit. If the renewal narrows disclosure further than Anthropic prefers, that creates tension with the lab's stated values. The lab will probably sign anyway — the political cost of walking away is too high — but the terms could affect Anthropic's longer-term strategic positioning.
The OpenAI Specific Angle
OpenAI's position is also structurally different. The lab has had a more complex relationship with safety evaluation over the past 18 months, has been less publicly vocal about federal oversight, and has been more focused on capability advancement and commercial deployment. The OpenAI posture suggests the lab will sign a renewal if the terms are workable but is less likely to fight for more aggressive disclosure provisions than the May 5 template provides.
The interesting question for OpenAI is whether the renewal includes terms around the deployment of frontier-tier capabilities to enterprise customers. OpenAI's commercial trajectory under the GPT-5.5 release has been heavily focused on enterprise deployment, and the agreements could include provisions about how evaluated capabilities are deployed to specific customer categories. This is speculation, but the structural fit between OpenAI's commercial focus and possible agreement terms is suggestive.
The Q3 Timeline
The realistic window for the renegotiation to resolve is June through September 2026. This is the window between the May 5 Google/Microsoft/xAI signings and the next major OpenAI and Anthropic releases. Resolving inside this window allows both labs to ship the next generation of frontier models with the federal evaluation regime in place.
If the renegotiation slips past September without resolution, the announcement of next-gen models becomes the forcing function. Both labs would face the question: ship without federal evaluation (Outcome 2 by default) or delay release until agreements are in place (which carries commercial cost).
The base case is that announcements happen in late Q2 or Q3, with agreement structure similar to May 5 and possibly with minor custom terms for each lab. The next Opus release and the next major OpenAI release both flow through the new evaluation process. The federal evaluation regime achieves full coverage of US frontier labs by Q4 2026.
The downside scenarios are non-zero but unlikely. We watch the announcements as they happen.