OpenAI's $500 ChatGPT Pro Max Tier Sells Compute Priority, Not More Compute

On September 24, strings buried in ChatGPT's unreleased frontend code and later confirmed in OpenAI's public Codex repository revealed that the company is preparing a subscription tier called ChatGPT Pro Max, priced at $500 a month. Its only stated differentiator from the existing $200 Pro plan is access to what the code labels "Fastest Work and Codex." OpenAI has made no official announcement and set no launch date, but the evidence is real — PR #47971, merged into OpenAI's public Codex repository on September 25, confirms the strings. The leak surfaced five days before OpenAI DevDay on September 29.

What the leak reveals is as important for what it does not say as for what it does. Pro Max does not promise more usage quota, more models, or capabilities unavailable on lower tiers. The pitch is speed — specifically the speed at which multi-step agentic tasks execute. For the class of users who run ChatGPT Work or Codex continuously across a working day, that distinction redefines what an AI subscription is actually selling.

Why Inference Speed Compounds Across Agentic Workloads

ChatGPT Work is OpenAI's long-running agent mode, built for tasks measured in hours rather than a single exchange. Codex is OpenAI's coding agent, capable of operating on a project for hours or days. Both share a dedicated compute pool separate from standard chat traffic. Each step an agentic system takes requires its own model call: reading files, writing code, running tests, interpreting output, retrying. A complex Codex session might generate hundreds of such calls.

At fourteen times the standard inference speed — the throughput OpenAI demonstrated with its Cerebras-powered Ultrafast mode in August — a session that would otherwise take hours completes in a fraction of that time. For developers billing clients by the project, or research teams on competitive schedules, halving turnaround on a Codex job is worth the premium. That is what Pro Max appears to be selling: not more compute, but your job ahead of the queue.

Cerebras: The Infrastructure Behind "Fastest"

OpenAI's relationship with Cerebras, whose wafer-scale processors hold the largest on-chip memory of any AI accelerator, goes back to early 2026. In February, OpenAI deployed GPT-5.3-Codex-Spark on Cerebras hardware as a research preview for Pro users — its first production model on that infrastructure — running at more than 1,000 output tokens per second, roughly fifteen times faster than the Codex version it replaced.

In August, OpenAI previewed GPT-5.6 Sol Ultrafast at 750 tokens per second — fourteen times standard processing — in a limited API preview. Cerebras's own benchmarking found a 5.6-times end-to-end speedup on knowledge-work tasks with no measurable quality loss. Neither OpenAI nor Cerebras has confirmed a Cerebras link for Pro Max; TestingCatalog described it as "plausible." The structural inference is straightforward: the only speed-priority commercial tier OpenAI has built so far runs on Cerebras hardware, and Pro Max is a speed-priority product.

Why This Tier Exists: Astra Exhausted the $200 Plan's Economics

The immediate cause of Pro Max is GPT-6 Astra, launched September 3. Astra introduced direct computer use and browser autonomy — capabilities that consume substantially more inference resources than conventional text generation, because autonomous task completion requires the model to execute dozens or hundreds of operations per session.

One week after launch, on September 10, OpenAI's head of core products Thibault Sottiaux announced a pause on new $200 Pro subscriptions, citing the tier as the source of "greatest strain" on OpenAI's systems. As of September 25, that pause remains in effect with no restoration timeline. The economic problem is structural: the $200 flat rate was priced before Astra's compute demands were known. Heavy Pro users running hours-long Codex sessions extracted far more server time per dollar than the plan's average user assumed. Pro Max, priced closer to what sustained agentic compute actually costs, resolves the subsidy.

Read more: OpenAI's GPT-6 Astra launch and unprecedented compute demands

The Industry Divergence: Price vs. Identity

OpenAI's approach to allocating scarce frontier compute — charge more — is not the only path available to a large AI lab. Anthropic, which released Claude Opus 5.5 on September 22, routes requests involving offensive cybersecurity to Claude Opus 4.8 and flagged biology-related requests to Claude Opus 5 automatically, reserving full Opus 5.5 capability for users who have passed verification through its Life Sciences or Cyber Verification programs. The model also incorporates safeguards against distillation — attempts to extract its capabilities through its outputs route to more constrained models.

Claude Max, at $100 or $200 a month, has not introduced speed-priority pricing. Anthropic is managing frontier access through identity and use-case gating rather than price-priority queueing. Where OpenAI makes frontier speed available to anyone who can pay $500 a month, Anthropic restricts certain capabilities regardless of payment. Both strategies reflect underlying assumptions about where risk in frontier AI lies — and who should bear the cost of managing it.

Deflationary Commodity, Inflationary Frontier

The same week that Pro Max strings appeared in ChatGPT's code, OpenAI cut prices at the opposite end of the market. On September 22 — ninety minutes before Anthropic announced Opus 5.5 — OpenAI launched GPT-6 Sol and GPT-6 Luna, halving API costs compared to the previous GPT-5.6 generation. GPT-6 Sol now costs $2 per million input tokens and $10 per million output. GPT-6 Luna dropped to $0.10 per million input and $0.50 per million output, below every frontier-class API on the market. OpenAI attributed the cuts to improved caching and inference efficiency, and confirmed the new rates carry no expiration date.

The simultaneous movements reveal a deliberate market structure: commodity AI — routine coding, summarization, classification — is converging toward near-zero marginal cost. Frontier AI capable of autonomous hours-long execution is moving in the opposite direction. This two-speed dynamic is not unprecedented in technology markets, but the speed of the divergence — OpenAI moved both poles in the same week — signals that OpenAI has made a structural decision rather than a temporary pricing adjustment.

What Remains Unresolved Before DevDay

The code strings confirm a name, a price, and a description. They do not confirm usage limits above what the $200 Pro plan already provides — TestingCatalog explicitly noted that higher quotas "may also be part of the package, although that is not confirmed yet." They confirm no launch date, model identifier, or tier architecture. OpenAI's public pricing page still lists six plans with no Pro Max among them.

The precedent for this kind of leak is directly on point. Tibor Blaho spotted code strings for the $200 Pro plan in December 2024; OpenAI announced it publicly within hours. DevDay on September 29 is the earliest plausible official announcement window, though no launch is guaranteed. What is determinable from the evidence already available is that OpenAI has, for the first time, coded a consumer subscription tier whose primary differentiator is not how much of a model you receive but how quickly your work executes. The company will need to explain at DevDay — or eventually — whether that $500 price point reflects a temporary response to compute scarcity created by Astra's launch or a permanent architecture for how frontier AI gets allocated among those who want it most.