OpenAI's GPT-5.5 Retrain Delivers Speed and Smarts
OpenAI released GPT-5.5 (codenamed Spud) on April 23, 2026, one week after Anthropic shipped Opus 4.7. Unlike previous GPT-5 variants, this is a fully retrained base model with new weights and training data, not a fine-tuned refinement. Key improvements include native omnimodal processing (text, image, audio, video through unified architecture), matching GPT-5.4 latency while delivering higher intelligence, and significantly lower token consumption on coding tasks. The release includes a tiered lineup with a base variant and a higher-capability Pro tier targeting enterprise and specialized work.
TL;DR
- →GPT-5.5 is OpenAI's first fully retrained base model since GPT-4.5, not a point release or fine-tuned variant of GPT-5.4
- →Native omnimodal architecture processes text, images, audio, and video through a unified system rather than separate pipelines, enabling coherent cross-modal reasoning
- →Achieves GPT-5.4 latency while delivering measurably higher intelligence and using significantly fewer tokens for Codex tasks, addressing the speed-versus-capability tradeoff
- →Tiered release strategy includes base GPT-5.5 for general use and GPT-5.5 Pro at $30 per million input tokens and $180 per million output tokens for high-accuracy enterprise work
Why it matters
GPT-5.5 represents a meaningful architectural shift rather than incremental improvement, arriving just days after Anthropic's Opus 4.7 in what appears to be an accelerating release cadence at the frontier. The unified omnimodal design and training-level behavioral changes signal that both OpenAI and Anthropic are converging on agentic capabilities and multi-modal reasoning as the next competitive battleground. The latency-matched intelligence gain directly addresses a persistent pain point for production users watching costs climb without proportional capability gains.
Business relevance
For operators running OpenAI infrastructure, this release offers the first genuine cost-efficiency improvement since GPT-4, with lower token consumption and higher output quality at equivalent latency. The tiered pricing structure and Pro variant targeting enterprise and legal work suggest OpenAI is segmenting the market to capture higher-value use cases, while the unified omnimodal backbone enables new product categories like the announced super app fusing chat, coding, and browsing. Teams evaluating between OpenAI and Anthropic now face a narrower capability gap but sharper architectural differences in how each company approaches multimodal reasoning.
Key implications
- →Full retraining cycles are becoming the standard for major releases rather than fine-tuning, suggesting frontier labs are willing to absorb retraining costs to achieve qualitative behavioral shifts
- →Unified omnimodal architecture from training rather than bolted-on adapters may become table stakes for agentic models, raising the bar for smaller competitors relying on modular approaches
- →The speed-versus-capability tradeoff that has constrained production deployment may be genuinely resolved, unlocking new use cases in latency-sensitive applications like real-time coding and authentication flows
- →Tiered pricing with 6x multiplier for Pro tier suggests frontier labs are testing willingness to pay for specialized high-accuracy variants, potentially creating a bifurcated market
What to watch
Monitor whether GPT-5.5's latency-matched intelligence gain translates to measurable cost savings and capability improvements in production deployments over the next 30 to 60 days. Watch for Anthropic's response to the unified omnimodal architecture and whether they pursue similar retraining or defend their modular approach. Track adoption of the GPT-5.5 Pro tier and whether the 6x pricing multiplier signals a sustainable market for specialized high-accuracy models or represents overpricing that limits uptake.
ChatGPT 5.5 just landed. Here is what Spud actually brings to the fight.
OpenAI dropped its first fully retrained base model since GPT-4.5 exactly one week after Anthropic shipped Opus 4.7. The gap between these two flagships is smaller than the headlines suggest, and more interesting than either company is admitting.
Seven days. That is the gap between Claude Opus 4.7 landing on April 16 and GPT-5.5 shipping this morning under the codename Spud. Two frontier flagships, both pitched at agentic work, both running 1M-token context windows, both leaning on thinking-style reasoning. If you read the launch posts side by side, you would think the companies were racing to ship the same product. They are not. Spud is a fundamentally different animal than anything OpenAI has shipped since GPT-4.5, and it leaves the older GPT-5 family behind in ways that matter more than the spec sheet suggests. Here is what is actually in the box, and where it beats the model Anthropic shipped a week ago.
A full retrain, not a refinement
The first thing to understand about GPT-5.5 is that it is not a point release. It is the first fully retrained base model OpenAI has shipped since GPT-4.5, which means the weights underneath are new, the training data is new, and the behavior is not a tuned-up GPT-5.4 with a faster wrapper. Everything that preceded it in the GPT-5 family — the 5.0 launch in August 2025, the 5.4 variants, the Thinking mode iterations — was built on top of the same foundation. Spud is a different foundation.
That distinction matters because the qualitative changes people are reporting in early testing are not the kind of thing you get from fine-tuning. The model handles ambiguity differently. It sticks with hard problems longer. It uses fewer tokens to get to the same answer. Those are training-level behaviors, not prompt-level tricks.
Omnimodal from the ground up
GPT-5.5 is natively omnimodal. Text, images, audio, and video are processed through a unified system rather than bolted on through separate pipelines. Earlier GPT-5 models supported multimodal inputs, but the processing was not unified — you could feel the seams when a model reasoned about an image it had been handed by a vision adapter.
The practical effect is cross-modal reasoning that holds up. The model can look at a screenshot, listen to a voice memo describing what is wrong with it, and produce a coherent written plan that references both. This is also why Spud is the backbone OpenAI is using to power its upcoming "super app," which fuses chat, coding, and browsing into one surface.
It matches 5.4 latency at a higher intelligence level
This is the detail that should make infrastructure people stop scrolling. GPT-5.5 matches GPT-5.4's per-token latency in real-world serving while performing at a measurably higher level of intelligence. It also uses significantly fewer tokens to complete the same Codex tasks. That is the two-front win OpenAI has been promising and mostly failing to deliver since GPT-4 — smarter and faster, not smarter or faster.
For anyone who has been watching their OpenAI bills climb while waiting for a serious capability jump, this is the release where the curve finally bends the right way.
A tiered lineup that forces a choice
OpenAI split the release into variants that actually do different things, not just different prompts:
GPT-5.5 (base, with Thinking)
Rolling out today to Plus, Pro, Business, and Enterprise users in ChatGPT and Codex. This is the everyday workhorse — 1M context in the API, 400K in Codex, and the Thinking mode that lets it deeply self-check authentication flows and multi-file edits on the first pass.
GPT-5.5 Pro
A higher-capability variant targeting extreme-accuracy work, heavy-duty enterprise tasks, and legal research. API pricing lands at $30 per million input tokens and $180 per million output tokens, which is six times the base model and signals exactly who it is for.
Fast mode in Codex
A genuinely new lever. Codex's Fast mode generates tokens 1.5x faster for 2.5x the cost. For any developer who has watched an agent loop chew through a half-hour task, being able to buy your way out of latency at a predictable ratio is a real tool.
Operator Note
The base/Pro/Fast split means you can no longer pick "GPT-5.5" without making a routing decision. Base for most work, Pro for precision-critical runs, Fast when the wall clock matters more than the invoice.
The Cyber program nobody else has
Here is a launch detail that did not make most of the recap posts: GPT-5.5 is now capable enough at identifying and patching advanced security vulnerabilities that OpenAI shipped stricter cyber-risk classifiers for general users, paired with a new Trusted Access for Cyber program that grants a "cyber-permissive" license to verified security professionals.
OpenAI is also classifying the biological, chemical, and cybersecurity capabilities of GPT-5.5 as High under its Preparedness Framework. For anyone working in enterprise security, this is the first frontier model that comes with a recognized licensing path for offensive research. Anthropic's equivalent — Claude Mythos Preview — sits above Opus 4.7 on capability but is gated to defensive work only. Different philosophies, different doors.
Where the benchmarks actually diverge
The honest read across independent reviewers is that on the benchmarks both companies report, Opus 4.7 leads on six and GPT-5.5 leads on four. But the leads cluster by workload, not by overall quality. That is the frame worth keeping in mind before picking sides.
Left column: GPT-5.5. Right column: Claude Opus 4.7. Highlighted score indicates the winner on each benchmark.
The sleeper: long-context retrieval
Both models advertise 1M-token context windows. That is the number every comparison post cites. What almost nobody is mentioning is what happens at the upper end of the window — specifically, how reliably each model retrieves information placed deep inside a long context.
On OpenAI's MRCR v2 8-needle benchmark, the gap between GPT-5.5 and Opus 4.7 at the 512K to 1M range is the largest single discrepancy in the entire head-to-head. If you are building agents that routinely reason over entire codebases, full policy corpora, multi-document research packets, or hour-long agent traces, context-size parity does not mean retrieval parity. This is the kind of gap that changes architecture decisions, not prompt decisions.
Caveat
MRCR v2 is an OpenAI-reported benchmark. The directional claim matches everything else Spud is doing well, but this is one worth waiting for independent validation on before rebuilding your retrieval pipeline around it.
Terminal work: where Spud genuinely pulls ahead
Terminal-Bench 2.0 is the benchmark that captures long-running shell and CI work — exactly the kind of task a serious coding agent needs to hold together for hours without losing the thread. GPT-5.5 posts 82.7%. Opus 4.7 posts 69.4%. That is a thirteen-point gap on the benchmark that most closely resembles real production agent work.
The caveat, and it is a real one, is that OpenAI ran Terminal-Bench 2.0 with a specialized Codex CLI harness while Anthropic used the Terminus-2 scaffold. Harness choice matters. But even when you discount for the home-field advantage, GPT-5.5's lead on long-horizon terminal tasks is consistent enough across independent reviewers to take seriously.
Coding: Opus is still the repo-level model
Flip to SWE-Bench Pro — the multi-language, repo-scale coding benchmark that tests pull requests and architectural work rather than shell commands — and the picture inverts. Opus 4.7 lands at 64.3% versus GPT-5.5 at 58.6%. Anthropic's self-verification advantage shows up when the task is "refactor this system" rather than "automate this pipeline."
The cleanest mental model: GPT-5.5 runs the shell. Opus 4.7 writes the code. Both can do the other job, but the one they are each best at is clearly drawn in the data.
Vision, pricing, and latency
Vision
Opus 4.7 reads images at roughly 3.3x the resolution of any comparable model — up to 2,576 pixels on the long edge, about 3.75 megapixels. GPT-5.5's omnimodal architecture is broader, especially with audio and video, but for anything involving screenshots, diagrams, or financial charts, Opus still has the pixel advantage. For anyone whose agents live inside dense UIs, that matters more than unified modality.
Pricing
Both list at $5 per million input tokens. Output is where they part ways:
GPT-5.5 runs $30 per million output tokens. Opus 4.7 runs $25 per million output, but applies a 2x surcharge above 200K input tokens. Translation: Opus is cheaper on short-to-medium prompts, GPT-5.5 is cheaper once you cross the 200K input threshold. Batch and Flex pricing on the OpenAI side lands at half the standard rate, and Priority processing runs 2.5x the standard rate.
Latency
This one surprised me. Opus 4.7 has time-to-first-token around 0.5s. GPT-5.5 baseline sits near 3s. That is roughly six times faster to first token on Anthropic's side. For interactive applications where the user is watching the cursor, Opus feels more responsive. For batch work where the user is not watching, it does not matter.
The decision framework
Here is the practical routing logic based on everything above. This is what I would build into a production system today.
Send to GPT-5.5
Agents that drive a terminal or CI pipeline for hours. Browser agents and web research tasks. Computer-use work where ambiguity is the norm rather than the exception. Knowledge work across documents, spreadsheets, and analysis. Anything where the task runs over 500K tokens of context and retrieval reliability matters. Cybersecurity research if you have the trusted-access license.
Send to Opus 4.7
Repository-level refactors, pull requests, architectural work. Computer-use agents that read high-resolution screenshots or dense financial charts. Reasoning-heavy scientific and technical work. Interactive applications where time-to-first-token defines the user experience. MCP-heavy integrations where Anthropic's ecosystem has the advantage. Financial agent work. Any task where a wrong answer is worse than no answer — Opus admits when data is missing instead of filling the gap with a confident guess.
Use both
Honestly, this is where most serious teams will land. Route terminal and long-context work to Spud. Route code review, precision analysis, and pixel-dense vision to Opus. The cost of maintaining both providers is already baked into any production AI stack worth running. The cost of picking one and eating the failure modes of the other is much higher.
The bottom line
Spud is the first OpenAI release in a long time that reshapes what the frontier looks like rather than incrementing it. Full retrain, native omnimodal, matching latency at higher intelligence, a licensing program nobody else has, and a genuine lead on long-horizon terminal work and deep-context retrieval. It does not wipe out Opus 4.7, which is still the stronger model for repo-scale coding, precision reasoning, and anything that lives inside a high-resolution screen. But for the first time since the GPT-5 family launched last summer, OpenAI has shipped a model that forces Anthropic to respond on capability rather than on price. The next six months are going to be interesting.
References & Sources
- OpenAI. "Introducing GPT-5.5." openai.com (April 23, 2026). Primary source for full-retrain status, omnimodal architecture, latency and token-efficiency claims, API pricing, GDPval and Tau2-bench Telecom results, and the Trusted Access for Cyber program.
- Anthropic. "Introducing Claude Opus 4.7." anthropic.com/news/claude-opus-4-7 (April 16, 2026). Source for image resolution specs, Opus 4.7 positioning relative to Claude Mythos Preview, and SWE-Bench Pro results.
- VentureBeat. "GPT-5.5 launches with cyber-permissive licensing and state-of-the-art benchmarks." venturebeat.com (April 2026). Source for Trusted Access for Cyber program details and OpenAI's claimed 14-benchmark lead.
- The New Stack. "GPT-5.5 Variants: Base, Pro, Thinking, Fast." thenewstack.io (April 2026). Source for variant breakdown and rollout tiers.
- Fortune. "OpenAI's GPT-5.5 is a big step toward agentic computing." fortune.com (April 2026). Source for agentic computing framing and Greg Brockman commentary.
- Seattle Times. "Inside OpenAI's GPT-5.5 launch." seattletimes.com (April 23, 2026). Source for Spud codename, coding capability claims, and rollout details.
- 9to5Mac. "GPT-5.5 Codex Fast mode pricing." 9to5mac.com (April 2026). Source for Codex Fast mode 1.5x speed and 2.5x cost ratios.
- Digital Applied. "GPT-5.5 vs Claude Opus 4.7: Benchmark Breakdown." digitalapplied.com (April 2026). Source for SWE-Bench Pro comparison, MCP Atlas scores, and MRCR v2 long-context retrieval gap.
- LLM Leaderboard. "Opus 4.7 vs GPT-5.5 aggregate comparison." llmleaderboard.com (April 2026). Source for aggregate benchmark leadership (6-4 split), CharXiv-R scores, latency numbers, and Opus 4.7 image resolution comparison.
- R&D World. "HLE and Terminal-Bench head-to-head." rdworldonline.com (April 2026). Source for HLE scores (with and without tools) and Terminal-Bench 2.0 harness differences.
- Yahoo Finance. "OpenAI positions GPT-5.5 for real computer work." finance.yahoo.com (April 2026). Source for OSWorld-Verified and BrowseComp results.
- MoneyControl. "GPT-5.5 omnimodal capabilities explained." moneycontrol.com (April 2026). Source for native omnimodal architecture details.
vff Briefing
Weekly signal. No noise. Built for founders, operators, and AI-curious professionals.
No spam. Unsubscribe any time.
