GPT-5.5 'Spud': OpenAI Retrains the Base Model and Starts Competing on Unit Economics

The version number undersells it. Most of the GPT-5.x releases over the past year, 5.1 through 5.4, were post-training layers on the same base model shipped in August 2025. GPT-5.5, launched April 23 in ChatGPT and Codex with API access on April 24, is different: a from-scratch retraining, new pretraining corpus, rewritten architecture, built around an agentic objective. Internally it was codenamed "Spud." It is the most substantial thing OpenAI has shipped since GPT-5.0, and the most revealing thing about it is not the model at all. It is the pricing philosophy underneath.

What it is built to do

OpenAI is refreshingly narrow about the target. GPT-5.5 is not pitched as a generalist intelligence leap; it is a deliberate specialization toward agentic, multi-tool, long-horizon work: coding, computer use, knowledge work, and early scientific research. Greg Brockman called it "a new class of intelligence" and "a big step toward more agentic and intuitive computing" on the press call, which is the sort of thing executives say, but the benchmark deltas back the direction if not the grandiosity.

The gains cluster exactly where OpenAI aimed them. On Terminal-Bench 2.0, a measure of navigating real shell environments, it jumped to 82.7% from GPT-5.4's 75.1%. Long-context coherence, the thing that actually breaks agentic pipelines, improved dramatically: MRCR at 1M tokens went from 36.6% to 74.0%. OpenAI frames three behaviors as the real difference from 5.4: holding context across large multi-file systems, recovering from ambiguous tool failures rather than just visible bugs, and checking its own work before declaring a task done. For anyone who has watched an agent confidently complete half a task and hallucinate the rest, self-verification is the feature that matters.

The benchmark honesty section

Here is where a fair write-up diverges from the launch post. "Strongest agentic coding model" is not the same as "best at every coding task," and OpenAI's own numbers show it. GPT-5.5 leads decisively on Terminal-Bench and long-context retrieval, but Anthropic's Claude Opus 4.7 still leads on SWE-Bench Pro (64.3% vs 58.6%), the more stringent test of resolving real issues in enterprise codebases, and on MCP-Atlas tool orchestration. The practical read from the reviewers is clean: GPT-5.5 is stronger when the task is "complete this terminal workflow," Opus is stronger when it is "fix this production bug." Gemini 3.1 Pro, meanwhile, holds its own on pure reasoning benchmarks neither OpenAI nor Anthropic tops.

One caveat worth flagging with appropriate skepticism, since it comes from a single evaluation rather than a consensus: at least one analysis reported a high out-of-domain hallucination rate on factual questions outside the model's tuned domains. If accurate, it is a reminder that a model optimized hard for agentic coding is not automatically the one you want drafting a legal memo. Treat that number as unconfirmed pending broader testing, but treat the underlying point, specialization has costs, as sound.

The pricing debate, and why it misses the point

The loudest reaction was to price. Standard GPT-5.5 lists at $5 per million input tokens and $30 per million output, with GPT-5.5 Pro at $30/$180. Depending on which prior model you anchor to, that is either a doubling (versus GPT-5.4's $2.50/$15) or roughly flat (versus GPT-5.3-class pricing). The "it doubled" framing went viral. It is also, for the actual use case, misleading.

The number that matters is cost per completed task, not cost per token. OpenAI claims, and early developer calculations corroborate, that GPT-5.5 uses roughly 40% fewer output tokens on Codex-style tasks, which pulls the effective cost increase for agentic pipelines closer to 20% than 100%. And it compounds: if a model completes more tasks correctly on the first pass, the tasks that would have needed a manual rewrite or a second round-trip disappear from the bill entirely. One cost analysis put it well: if one in eight tasks on the older model needed a manual fix, the newer, pricier model is cheaper in aggregate. For offline work, Batch pricing halves the rate back to exactly GPT-5.4's old numbers.

The strategy that actually changed

Strip away the model and the benchmarks and the real news is a posture shift. An analyst note the morning after the launch nailed it: OpenAI is now competing on unit economics, not just capability, a fundamentally different stance than its 2025 releases, and one that suggests the marginal cost of training and serving these systems has come down materially. The pricing is explicitly built to undercut Anthropic on the coding lane (GPT-5.5 lists below Opus 4.7 on output) while the token-efficiency story lets OpenAI raise the sticker price and still claim a lower bill.

Underneath sits Brockman's "super app" ambition: fusing ChatGPT, Codex, and an in-house AI browser into a single enterprise product, aimed at the same all-in-one vision others are chasing. The scale behind it is the point, 900 million weekly users, over 50 million subscribers, 4 million Codex developers, and 85% of OpenAI's own staff using Codex weekly. GPT-5.5 is the engine; the super app is the car.

The 30,000-foot read

The honest summary resists both the hype and the sticker shock. GPT-5.5 is a genuine base-model advance, narrowly and deliberately aimed at agentic work, where it now leads, while trailing on enterprise-codebase bug-fixing where Anthropic holds firm and on general factual reliability where specialization may have cost it something. The price "doubling" is mostly an accounting illusion once you measure cost per completed task rather than per token.

The deeper signal is the cadence and the framing. Six weeks between 5.4 and 5.5 is not a schedule, it is competitive pressure, and OpenAI's pivot from selling raw intelligence to selling cost-per-useful-task tells you where the frontier fight is heading. 2026 is no longer the year of getting the smartest model. It is the year of getting the cheapest reliable one, and the labs have all quietly agreed those are different problems.

Sources: Verdent guide (launch dates, four target areas, self-verification behaviors); TokenMix (benchmark deltas, "Spud" codename, availability); DDR Innova (Terminal-Bench vs Opus, SWE-Bench Pro caveat, long-context gains); Tech Insider (token-efficiency math, unit-economics analyst note); Apidog (full pricing, Batch/Flex tiers, cost-per-task worked example); CometAPI (competitor price snapshot); Pasquale Pillitteri (retrained-base-model detail, hallucination-rate caveat). All benchmark figures are OpenAI's own reported numbers unless noted; the 40%-fewer-tokens claim is OpenAI's, corroborated by early developer calculations. The high out-of-domain hallucination figure is a single-source evaluation and is flagged as unconfirmed. Opus 4.7's exact launch timing is covered in its own post; dates cited here for comparison come from secondary coverage.