GPT-5.5 'Spud' and the April 2026 Foundation Model Reset: Every Major Lab Doubled Token Pricing in Thirty Days

Four Frontier Labs, Thirty Days, One Coordinated Pricing Signal

In the thirty days between March 28 and April 27, 2026, every major frontier AI lab shipped a fundamentally agentic-first foundation model and every one of those models came with token economics that would have been considered uncompetitive twelve months ago. The compression of the release calendar is unusual. The convergence of the pricing actions is more unusual. Read together, the April 2026 cohort represents the clearest signal yet that the foundation model category has completed its transition from chat completion engine to action engine — and that the labs have priced the new unit of value at a premium reflecting the compute, memory, and supervision cost of running it.

The headline release is GPT-5.5. OpenAI shipped the model on April 23, 2026, internally codenamed 'Spud' — the first complete retraining the company has performed since the GPT-4.5 era and a clear architectural break from the incremental GPT-5 series that defined late 2025. The model immediately reclaimed the number one position on the Artificial Analysis Intelligence Index with a record score of 60 and posted a 78.7 percent score on OSWorld-Verified, the benchmark that measures whether a model can independently navigate operating systems and complete realistic multi-application tasks.

The pricing is what made the release news beyond the AI community. GPT-5.5 launched at five dollars per million input tokens and thirty dollars per million output tokens — exactly twice the cost of GPT-5.4 and roughly four times the cost of GPT-5 base. The price action runs counter to the inference-cost-trends-toward-zero narrative that defined frontier model economics through 2024 and 2025, and it is not isolated.

The Other Three Releases

Anthropic shipped Claude Mythos 5 on April 25, two days after GPT-5.5. The ten-trillion-parameter model is positioned for high-stakes cybersecurity and coding workloads and entered limited release through a partner-only distribution channel. Anthropic has declined to publish standard API pricing but partner sources place the enterprise tier within ten percent of GPT-5.5's input and output pricing. The model card emphasizes long-horizon coding capability, with notable gains on SWE-Bench-Live and AgentBench-Production — benchmarks that measure realistic codebase-wide work rather than the snippet completion problems that defined HumanEval.

Google DeepMind shipped Gemini 3.1 on March 28. The release added real-time voice and image streaming, persistent multi-day context, and a six-times memory compression that allowed Google to keep headline pricing flat while the underlying compute load on agent-class workflows grew. The pricing decision is itself a strategic signal — Google is willing to absorb margin compression in exchange for setting the default interface convention before the rest of the industry. Gemini 3.1's effective per-completed-task cost is higher than its predecessor's because the average completed task now consumes seven to ten times more tokens, but the headline price chart looks unchanged.

DeepSeek shipped V4 on April 27, the only release in the cohort to ship as open weights. The model's architecture assumes long-horizon agent loops as the default execution mode rather than a layer bolted on top of a chat-completion substrate. The hosted API price actually fell relative to DeepSeek V3 — the only price decrease in the April cohort — and the combination of open-source distribution, native MCP tool integration, and aggressive pricing has already prompted enterprise coding tools including Cursor, Replit, and GitHub Copilot Workspace to quietly update their enterprise tier pricing in response.

Why the Math Forced Higher Prices

The pricing actions are not coordinated in the antitrust sense, but they are coordinated in the structural sense — every lab independently arrived at a similar conclusion about what an agent-era unit economy looks like. A chat completion has a roughly known cost structure: a few thousand input tokens, a few hundred output tokens, one inference call, no tool roundtrips. An agent task has a fundamentally different structure. The model plans a sequence of steps, calls tools, observes outputs, replans, retries failed steps, manages working memory that grows over the task, and produces a final result that may have required dozens of inference calls and tens of thousands of context tokens.

By the time an agent is performing a one-hour long-horizon task, it has consumed roughly two orders of magnitude more tokens than a chat completion would have. By the time an agent is operating a full computer-use session with screenshots, GUI interactions, and persistent state, it has consumed three orders of magnitude more. At chat-era pricing, the median completed coding agent task on a real production codebase was running between thirty and seventy cents on the largest frontier models. At GPT-5.5 pricing the median completed coding agent task crosses one dollar, and long-horizon variants run several dollars per task.

The labs are pricing for the workflow, not the call. They are charging more per token because the average call now sits inside a workflow that consumes many more tokens. The headline doubling of input and output prices is explicit recognition that the chat-completion price unit is no longer the correct denominator. The new denominator is the completed agent task.

The Benchmark Migration That Already Happened

The benchmarks that frontier model releases now emphasize have shifted faster than the public narrative has caught up to. MMLU, HumanEval, GSM8K, and BIG-Bench-Hard — the benchmark suite that defined the chat era — collectively occupied around ninety percent of model card real estate in 2023. By 2026 they collectively occupy under thirty percent. The gap has been filled by benchmarks that test fundamentally different capabilities: OSWorld-Verified (operating-system control), SWE-Bench-Live (codebase-wide refactoring), AgentBench-Production (long-horizon agentic workflows), and GAIA (multi-step reasoning over real artifacts).

The migration matters because benchmark attention drives training. A benchmark that everyone evaluates on becomes a benchmark that everyone tunes for, and a benchmark that everyone tunes for shapes the behaviors a model emits in production. The chat-era benchmarks rewarded fluent, confident prose responses to bounded questions. The agent-era benchmarks reward calibrated plans, reliable tool calls, accurate intermediate observations, and clean recovery behavior under realistic failure modes.

What the Reset Breaks for Engineering Organizations

Three architectural assumptions that defined enterprise AI deployments between mid-2023 and late 2025 are now broken. The assumption that per-call inference cost trends to zero has reversed at the agentic frontier — the per-task cost of a production agent has approximately doubled in twelve months. The assumption that retrieval-augmented chat is a sufficient substrate for serious agentic work no longer holds; production agents need durable working memory, tool orchestration, plan management, and evaluation harnesses that operate on real artifacts. The assumption that benchmark scores in published model cards are sufficient for model selection has become actively misleading because public benchmarks no longer track the workloads that matter for enterprise agentic deployment.

The deepest consequence is that the gap between organizations that can deploy agents well and organizations that cannot is widening, not narrowing. The chat era was strongly egalitarian — a small startup could run GPT-4 over an MVP for a few hundred dollars a month and reach parity with what an enterprise was doing. The agent era is structurally different. A serious agent deployment requires not just inference budget but evaluation infrastructure, observability tooling, supervision processes, and the engineering discipline to build the substrate around the model.

What to Watch Through Q3 2026

Three signals will tell us whether the April 2026 reset is the start of a durable agentic pricing regime or a transient overshoot before the next round of compression. Whether GPT-5.5 and Claude Mythos 5 hold pricing through Q3 2026 will signal whether the new tier is a floor or a temporary equilibrium. Whether DeepSeek V4's open-source closure rate stays under twelve months on production agentic capability will determine whether proprietary labs retain pricing power or face the same compression that hit the chat tier through 2024 and 2025. And whether private evaluation harnesses become standard inside Fortune 500 engineering organizations will determine whether public benchmark games continue to drive selection or whether real-workload measurement becomes the dominant decision input.

The labs went first because they always go first. The market follows. By the time the chat-era benchmarks have rotated entirely off the model cards — which on the current trajectory will be sometime in late 2026 — the foundation model category will have completed its transition from completion engine to action engine. The thirty days between March 28 and April 27, 2026, are the inflection point where the transition stopped being theoretical and started being priced.

For deeper analysis of the agentic foundation model reset, see our full analysis of the April 2026 reset and its implications for engineering organizations, which walks through the pricing math, the benchmark migration, the per-lab architectural bets, and the three-tier playbook for enterprise AI deployment in the new model generation.