There is a quiet agreement in how most people talk about the AI race. OpenAI versus Anthropic. ChatGPT versus Claude. Model versus model. Numbers versus numbers. But that framing misses the infrastructure sitting underneath all of it — the chips, the racks, the power draws measured in gigawatts, the hardware companies that every single AI lab depends on right now.
NVIDIA is not a neutral supplier in this story. It is, at this point, closer to the ground these companies are standing on. Both OpenAI and Anthropic train and serve their models on NVIDIA hardware. Both struck major deals with NVIDIA in the past six months. And both are now racing each other on the same silicon — which makes April 2026 a genuinely strange month to watch.
So let’s actually look at what happened. GPT-5.5 dropped on April 23, one week after Anthropic released Claude Opus 4.7. Same context window. Same general pricing tier. Both calling themselves the best agentic model out there. And underneath both of them: NVIDIA Blackwell GPUs, quietly doing the work.
What NVIDIA’s Role Actually Looks Like
Most people know NVIDIA makes GPUs. Fewer people think about what that means when you are training a model with hundreds of billions of parameters. The GB200 NVL72 — the rack-scale system that both GPT-5.5 and recent Claude models run on — connects 72 Blackwell GPUs into a single NVLink domain with 1.8 terabytes per second of GPU-to-GPU bandwidth. Before this architecture existed, the limit was eight GPUs connected per domain. That is a significant jump, and it is the direct reason why frontier models can now scale the way they do.
GPT-5.5 was specifically co-designed with the GB200 and GB300 NVL72 systems. OpenAI’s launch page says so directly. More interestingly, GPT-5.5 used Codex — its own agentic coding tool — to rewrite the infrastructure management software that runs on those NVIDIA racks, and got a 20% improvement in token generation speed out of it. So the model trained itself to run more efficiently on the hardware it was trained on. That is a bit recursive. Whether it is genuinely impressive or just a clever marketing line in the announcement, the hardware-software feedback loop is real.
On the Anthropic side, the picture is different but moving in the same direction. In November 2025, Microsoft, Anthropic, and NVIDIA announced a major three-way partnership. Anthropic committed to buying $30 billion of Azure compute capacity, with an option to scale up to one gigawatt of additional power. NVIDIA and Microsoft are investing up to $10 billion and $5 billion respectively in Anthropic. The deal also includes what NVIDIA called a “deep technology partnership” — Anthropic and NVIDIA engineers co-designing future Claude models for NVIDIA architectures, and vice versa.
So both companies are now in explicit hardware-software co-design relationships with NVIDIA. This was not the case two years ago.
GPT-5.5: What It Actually Does
GPT-5.5 shipped April 23, 2026, about six weeks after GPT-5.4 — which itself came out in March. The pace of releases has become a product story of its own. OpenAI’s chief scientist Jakub Pachocki said during the launch briefing that “significant improvements in the short term, extremely significant improvements in the medium term” should be expected to continue. That is a confident thing to say publicly.
The model has one central claim: it works like an agent, not a chatbot. You give it a messy task, and it figures out what to do next. That is the headline. Greg Brockman, OpenAI’s president, called it “a new class of intelligence” and said it feels like setting the foundation for how computer work will get done going forward.
The benchmark that keeps coming up is OSWorld-Verified, at 78.7%. This tests whether a model can operate real computer environments — not just answer questions about software, but actually use software. Open an application, navigate interfaces, complete tasks the way a person would. OpenAI no longer positions computer use as a demo feature. It is built into the product now, and GPT-5.5 is the first model where that seems genuinely true at scale.
Token efficiency is the other big one. GPT-5.5 uses 40% fewer tokens per Codex task compared to GPT-5.4. On long-context retrieval, the jump is even more dramatic — the MRCR v2 benchmark at one million tokens went from 36.6% on GPT-5.4 to 74.0% on GPT-5.5. That is more than double. For any developer running agents over large codebases or document sets, that number matters more than most others in the launch announcement.
And then there is the pricing. The API doubled. GPT-5.4 was $2.50 per million input tokens and $15 per million output. GPT-5.5 is $5 and $30. OpenAI’s argument is that the token efficiency more or less offsets this — you pay twice per token but use 40% fewer tokens per task, so the effective cost increase is around 20% for most coding workflows. That math holds if the efficiency gains apply to your actual workload. If they do not, you are just paying more.
The criticism that followed the launch is worth noting. Several independent reviewers pointed out that OpenAI’s system card includes an asterisk on SWE-bench Pro scores, noting “evidence of memorization” from other labs. Anthropic published a filter re-score analysis showing their Opus 4.7 margin holds on decontaminated subsets. OpenAI did not publish a matched re-run. This is not a minor footnote — benchmark contamination has been a known problem across the industry, and calling it out on a launch document is unusual. It has not stopped enterprise adoption, but the skepticism from developers tracking this carefully is real.
Where Claude Opus 4.7 Stands
Anthropic’s response to all of this is Claude Opus 4.7, which launched on April 16 — one week before GPT-5.5. The timing, in retrospect, was not accidental.
Opus 4.7 is built around a different priority. Where GPT-5.5 went after agentic workflows and computer use broadly, Anthropic went deep on coding precision. On SWE-bench Pro — the harder version of the coding benchmark where frontier models cluster around 50 to 65% — Opus 4.7 scores 64.3%. GPT-5.5 scores 58.6%. That 5.7-point gap on the harder benchmark is consistent with what earlier Opus releases did against OpenAI’s models. On the easier SWE-bench Verified version, GPT-5.5 edges ahead at 88.7% versus Opus 4.7’s 87.6%. So the question of which model is better at coding depends entirely on which version of the benchmark you are using, and which version better reflects your real workload.
For writing, the gap is more consistent. A systematic evaluation run by Dan Shipper’s publication Every found Claude Opus 4.5 scored 80% in writing quality versus GPT-5.2’s 74%. The comparison has not been formally repeated for 4.7 versus 5.5 yet, but based on what developers report anecdotally, the gap has not closed. If anything, Anthropic has leaned into this. Writing, instruction-following, and the ability to stay coherent across a long coding session are the things Opus is consistently praised for.
The pricing math is also different. Opus 4.7 is $5 per million input tokens and $25 per million output, versus GPT-5.5’s $30 on output. That 17% cheaper output matters for high-volume applications. But Anthropic changed their tokenizer with 4.7, and the new tokenizer uses 10 to 35% more tokens per input depending on content type — more for code, less for plain English. So the list-price advantage can disappear or reverse depending on what you are actually sending.
There is something Anthropic has that sits above Opus 4.7, though. Claude Mythos Preview, which launched in early April, is a model restricted to defensive cybersecurity workflows and available only to a small group of organizations. Anthropic said it found “thousands” of previously unknown software bugs using it. NVIDIA, Microsoft, and other critical infrastructure companies are among the organizations with access. GPT-5.5 does not have a direct equivalent — OpenAI has cybersecurity capabilities, but the tiered access model with a locked-down frontier version is something Anthropic moved on first.
Who Is Actually Winning
The honest answer: nobody is winning cleanly, and the question may be wrong. GPT-5.5 leads on terminal operation, long autonomous runs, and anything that involves driving software. Opus 4.7 leads on complex code reasoning, writing, and the harder benchmarks. Both have 1M-token context windows. Both cost roughly the same. Both run on NVIDIA hardware.
But something interesting happened this week among the developers paying closest attention. Several noticed that Opus 4.7 and GPT-5.5 feel like they swapped personalities relative to earlier versions. Opus 4.7 is more terse and contract-style in its responses — very precise plans, explicit scope, less conversational warmth. GPT-5.5 is more fluid and adaptive, closer to how older Claude versions felt. One newsletter called it “Uno reverse day.” That personality drift is real, and it affects which model people enjoy working with day to day, independent of the benchmark scores.
Dan Shipper, who runs the publication Every, figured out something that has been shared around a lot since yesterday. Use Opus 4.7 to write the plan, then hand that plan to GPT-5.5 to execute. On Every’s internal Senior Engineer benchmark, that combination scored 62.5 out of 100. Opus 4.7 alone scored in the low 30s. GPT-5.5 alone scored in the low-to-mid 40s. Human senior engineers score 80 to 90. So even the best single model is still significantly behind, but the combination approach almost doubles the individual model’s score. The workflow is not obvious — it depends on Opus 4.7’s tendency to write very specific, constraints-heavy plans, which GPT-5.5 can execute without needing further guidance.
This combination being the current best approach should probably be more surprising than it is.
The Hardware Dependency and What Might Change
Both OpenAI and Anthropic are deeply dependent on NVIDIA right now. OpenAI trains everything on Azure’s NVIDIA infrastructure. Anthropic committed to Grace Blackwell and Vera Rubin systems for its upcoming training runs, while also maintaining access to Google TPUs and AWS Trainium chips.
But Anthropic is reportedly exploring custom silicon. The company has not finalized any design or assigned a dedicated team, but the inquiry is happening. Developing an advanced AI chip reportedly costs around $500 million to get started. Anthropic also signed a long-term deal with Broadcom and Google for approximately 3.5 gigawatts of compute starting in 2027, when their new chips come online. So the dependency on NVIDIA is real today, but Anthropic is building a longer-term supply chain that does not rely entirely on a single hardware source.
OpenAI, for its part, is still on Azure and NVIDIA at scale. The co-design relationship between GPT-5.5 and the GB200/GB300 systems suggests they are going deeper into that dependency, not sideways. Whether that is a strategic constraint or a deliberate bet on NVIDIA’s continued hardware leadership is a question that will look clearer in 18 months.
NVIDIA’s Jensen Huang sent a company-wide email to NVIDIA’s 10,000-plus employees urging them to use GPT-5.5-powered Codex. He called it “lightspeed.” That is both a genuine endorsement and a business relationship functioning as expected — NVIDIA employees testing the model that runs on NVIDIA’s own GB200 NVL72 rack systems. The circular nature of it is not hidden. It is just the shape of how infrastructure and AI development have gotten tangled together.
What to Watch From Here
OpenAI is pushing toward what Greg Brockman described as a “super app” — combining ChatGPT, Codex, and an AI browser into a single service. The workspace agents feature launched April 22, one day before GPT-5.5. That timing was intentional: agents that can run long-horizon workflows, use connected apps, remember what they learned, and keep working when the user is away. The “super app” framing is still vague, but the product direction is clear. OpenAI wants one service that replaces the workflow of switching between tools.
Anthropic is doing something different. The Claude Partner Network, launched in Q1 2026, is a $100 million program that puts Anthropic engineers directly inside partner companies to build integrations. Dedicated account managers. Custom model fine-tuning on proprietary data. SOC 2 compliance support. It is not going after consumer attention — it is going after enterprise infrastructure deals.
API access to GPT-5.5 is not live yet. OpenAI says it requires additional cybersecurity safeguards before the API opens to developers. Nobody knows exactly when that will be. Anthropic’s Opus 4.7 API was available on day one across Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. For developers who need to build with the model right now, that availability difference is not trivial.
The April 2026 releases — GPT-5.5 on April 23, Opus 4.7 on April 16, both trained on NVIDIA infrastructure, both chasing enterprise contracts, both claiming agentic leadership — are not a moment where one lab pulled clearly ahead. They are a moment where the nature of what these labs are competing over got more specific. It is no longer “which model is smarter.” It is which product can actually run your workflow, on what cloud, at what cost, with what reliability. That question is still genuinely open.
The Benchmark Problem That Nobody Has Solved
There is a thing that keeps coming up in developer communities that the mainstream coverage mostly skips. Benchmarks are not neutral. They are not independent. The labs run their own evaluations, pick the benchmarks where their model scores best, and publish those. OpenAI led the launch announcement with Terminal-Bench 2.0 and GDPval. Anthropic leads with SWE-bench Pro. Both are defensible choices. Neither is the complete picture.
The MRCR v2 score that OpenAI highlighted — 74.0% at one million tokens — is impressive if accurate. But several developers noted that OpenAI did not explain its evaluation methodology in detail, and Anthropic has not published a comparable score for Opus 4.7 on the same benchmark. So the comparison is, technically, incomplete. That does not mean GPT-5.5 does not perform better on long-context retrieval. It probably does. But by how much, on your actual workload, versus Anthropic’s tokenizer-adjusted pricing? Nobody has an honest answer to that yet.
The contamination issue with SWE-bench Pro is separate from this but connected. If training data leaks into benchmark test sets — and there is now documented evidence that this has happened across multiple labs — then scores on those benchmarks measure something different than what they claim to measure. Anthropic re-scored Opus 4.7 on a decontaminated subset and said the margin held. OpenAI did not do the same with GPT-5.5. That is not an accusation. It may just be a difference in how each lab handles transparency around evaluations. But it is the kind of detail that matters when procurement decisions run into hundreds of thousands of dollars a month.
The hallucination numbers are also still a real issue across both models, just not discussed as much in April 2026 because the narrative has shifted to “agentic” capabilities. OpenAI’s own documentation on GPT-5.2 from December 2025 showed a hallucination rate of around 10.9% on thinking mode, down from 16.8% on GPT-5 thinking. Independent testing by Vectara at that time found slightly worse numbers than what OpenAI reported. GPT-5.5 claims further reductions, but independent re-verification has not caught up to the launch yet. The Claude models have historically tested better on factual reliability in writing contexts, which is part of why Anthropic has a stronger position in journalism, legal tech, and enterprise document workflows.
What This Actually Means for Someone Deciding Right Now
If you are building an agentic coding pipeline and need it working this week: Claude Opus 4.7 API is live across Bedrock, Vertex, and Azure Foundry. GPT-5.5 API is not yet available. That is the most concrete decision factor right now.
If long-context retrieval at 500K to 1M tokens is your main constraint: GPT-5.5’s numbers are more published and more consistent. Opus 4.7 has a 1M context window but has not published direct retrieval scores at that range.
If you are writing heavy: Opus 4.7. The gap in writing quality has held across versions and independent tests, and there is no evidence from the GPT-5.5 launch that this changed.
If you care about model behavior with ambiguous instructions: both models have improved, but GPT-5.5 explicitly claims better performance on interpreting unclear prompts without needing hand-holding. That claim is consistent with what early testers report, but “early testers” in this context means a subset of NVIDIA employees and enterprise partners who had pre-release access. Their feedback may not generalize.
And if you are planning a larger commitment — annual enterprise contracts, infrastructure decisions, the kind of thing that locks you in for 18 months — neither model has enough of a post-launch track record to justify betting entirely on it over the other. The multi-model routing approach that several developers are now recommending is probably the right default: Opus 4.7 for planning and complex reasoning, GPT-5.5 for execution and terminal-driven work, something cheaper for everything else.