David Sacks Rejects AI Pacing Rules: Pricing Impact

White House AI adviser David Sacks says OpenAI and Anthropic can slow their most advanced model development without waiting for new regulation. In a September 13 post on X, Sacks told the labs to “go ahead and pace the frontier,” but rejected the idea that they need regulatory approval, an antitrust waiver, or a shared framework to make that choice.

This is a policy dispute, not a model launch or price change. No provider announced a new model, rate, access policy, or retirement date. The live table, chart, and calculator above therefore remain the only defensible cost baseline.

What Sacks said

Sacks described OpenAI and Anthropic as a “duopoly on frontier intelligence” based on market share, revenue growth, and model capability. If their unreleased models appear too risky, he said, each company should slow voluntarily and accept the commercial tradeoff.

He argued that product-liability exposure and customer demand already push labs toward reliability and predictable behavior. He also accused the companies of using safety concerns to seek regulatory advantages over competitors and challenged the independence of model evaluator METR. Those are Sacks’s claims; his post did not provide evidence resolving METR’s governance or the legal reach of product liability.

Issue	Before Sacks’s post	After Sacks’s post
Frontier development	OpenAI and Anthropic publicly backed pacing	Their stated commitments remain unchanged
Regulatory path	Amodei proposed government-enabled coordination	Sacks publicly rejected regulation as a prerequisite
Independent evaluation	Anthropic committed to embedded evaluators; OpenAI said it would follow	No implementation detail or deadline changed
Public API pricing	Existing rate cards	Unchanged
Buyer action	Budget against released models	Keep doing so; no unreleased model is a purchasable SKU

What OpenAI and Anthropic proposed

Anthropic CEO Dario Amodei’s September 12 essay says labs should slow capability gains enough for safeguards to catch up. His three-stage framework begins with embedded third-party evaluators who receive employee-like access. Anthropic committed to that step immediately.

The second stage calls for frontier labs in democratic countries to coordinate common safety standards and limits on unchecked progress. Amodei said U.S. government mediation or a narrow antitrust waiver could enable those discussions. The third stage seeks narrower global coordination, including with China, around dangerous uses.

OpenAI CEO Sam Altman replied that he agreed with pacing and called embedded evaluators a good idea. He said OpenAI would adopt the practice and share more later. Neither company announced a training pause, release schedule, or rate-card change.

Pricing impact: stability versus market power

If voluntary pacing produces longer model lifecycles, buyers could benefit from fewer migrations, longer evaluation windows, and less prompt or toolchain rework. Reliability improvements could also lower cost per accepted task even if token rates stay flat.

The opposite risk is weaker competition. If leading labs coordinate release timing or safety thresholds, incumbents could preserve premium pricing and make compliance harder for smaller rivals. Government requirements could add audit and evaluation costs that eventually reach customers, although no company has quantified such costs.

Sacks’s preferred route avoids a new approval layer but leaves each lab to define what “pacing” means. Buyers cannot budget from that word alone. Compare current OpenAI pricing, Anthropic pricing, and Google AI pricing in the token calculator, then measure retries, latency, review time, and failure cost.

Official pages checked September 13 still list Claude Sonnet 5 at $2 input / $0.20 cached input / $10 output and Claude Opus 5 at $5 / $0.50 / $25 per million tokens. OpenAI still lists GPT-5.6 Sol at $4 / $0.40 / $20 and GPT-6 Astra Standard at $10 / $1 / $50 for short-context input, cached input, and output. These maintained rates feed the live comparison above; no regulatory or evaluator surcharge was published.

Labs decision: governance is not a model route

Sacks’s post and the evaluator commitments do not enter the Cost-per-Task Leaderboard. They provide no new model artifact, callable endpoint, fixed task set, token ledger, latency record, acceptance grader, or billed run.

We recorded an explicit governance and availability blocker. If either lab later identifies a paced release, changes access, or publishes a callable replacement, the exact route can be repriced or rerun instead of relabeling an existing result.

Who benefits—and who loses

Enterprise buyers benefit if pacing makes releases more reliable and gives security teams time to validate new models. Developers benefit if stable model IDs and longer deprecation windows reduce forced migrations.

OpenAI and Anthropic could benefit from longer monetization windows for current frontier models. They also risk losing customers to Google, DeepSeek, or open-model hosts if rivals keep advancing while they slow.

Smaller labs benefit from avoiding a regulatory structure designed around incumbent resources. They lose if voluntary coordination becomes an informal gate that limits access to evaluations, compute, or distribution.

What API buyers should do now

1. Keep budgets tied to released model IDs and published rates; ignore speculative prices for unreleased systems.
2. Ask vendors for deprecation notice periods, model-change controls, incident reporting, and fallback terms.
3. Maintain a second-provider route before any pacing policy changes release schedules or access.
4. Evaluate reliability as a cost input: track retries, human review, failed tool calls, and accepted outcomes.
5. Watch for concrete implementation details from OpenAI, Anthropic, and embedded evaluators—not political labels.

Our OpenAI vs Anthropic pricing guide compares the current buying decision. For the related debate over open-weight competition, read Garry Tan’s U.S. distillation proposal.

Teams seeking a managed alternative can compare Novita’s hosted open-model catalog. Verify the model, license, region, context limit, and current rate before switching production traffic.

Affiliate disclosure: AI Pricing Guru may earn a commission from the sponsored Novita link at no extra cost to you. It does not affect this analysis.

Bottom line

Sacks accepts the core action—slower frontier development—but rejects the proposed regulatory mechanism. OpenAI and Anthropic say pacing and embedded evaluation are needed; Sacks says they can make those choices now and let liability and customers discipline the market.

For buyers, nothing billable changed. Continue using live rates, demand clearer release and deprecation policies, and treat any future pacing framework as real only when it changes a model, contract, access rule, or invoice.

Sources: David Sacks’s full post on X, Dario Amodei’s “We Must Pace the Frontier” essay, Sam Altman’s response on X, the official Sanders–Casar proposal, the official Anthropic and OpenAI pricing pages, and AI Pricing Guru’s live pricing dataset. Sources checked September 13, 2026.