In six weeks this winter, all three frontier labs shipped their flagship models.
Claude Opus 4.6 dropped February 5. Gemini 3.1 Pro hit public preview February 19. GPT-5.4 landed March 5. By mid-March, the entire frontier had refreshed.
And no one won.
Interactive
The March 2026 Frontier
Three flagships launched within weeks. No clear winner. Select a model, then scroll down to see the benchmark split.
Context
272K tokens
Input cost
$2.50 / 1M tokens
Output cost
$15.00 / 1M tokens
Where it leads
General reasoning (ARC-AGI 2)
Computer use automation (OSWorld)
Professional knowledge work (GDPval)
Command-line tasks (Terminal-Bench)
Where it trails
Smaller context than Gemini
More expensive than Gemini at volume
Behind Claude on coding benchmarks
Benchmark category leaders
That is the most interesting result of March 2026. Not which model is best. The fact that the question is getting harder to answer.
The benchmarks split evenly
GPT-5.4 and Gemini 3.1 Pro score identically — 57 — on the Artificial Analysis Intelligence Index. Claude Opus 4.6 leads decisively on coding benchmarks. Gemini leads on mathematical and scientific reasoning. GPT-5.4 leads on general reasoning and professional tasks. Each model dominates different categories. None dominates all of them.
The spread on most individual benchmarks is 2–3 percentage points. That is within the noise for most practical applications.
This is not a temporary tie. It is a structural pattern. The frontier labs are converging.
Why convergence is the real story
The AI narrative for the last few years has been a capability race. Which lab is ahead. How far ahead. When the next leap comes. The assumption underneath all of it is that raw intelligence is the primary axis of competition.
Convergence breaks that assumption.
When three models launch within weeks and split benchmark categories with no clear winner, the competition is no longer mainly about intelligence. The models are smart enough. All of them. The question shifts to everything around the model: how much it costs, how fast it responds, how much context it can hold, how easily it integrates, and what you can actually build on top of it.
That is a fundamentally different competitive landscape.
The pricing tells the real story
Look at what the market is actually pricing:
Gemini 3.1 Pro offers a 2-million token context window at $2.00 per million input tokens. Claude Opus 4.6 offers 200K context at $5.00 per million input tokens. GPT-5.4 offers 272K context at $2.50 per million input tokens.
10x
Context window gap
Gemini 3.1 Pro offers 2 million tokens of context — roughly 10x what Claude Opus 4.6 offers. At equivalent tasks, you can run 5–7x more Gemini requests for the same cost.
That is not a capability story. That is an economics story. The models are converging on intelligence while diverging on delivery: different price points, different context sizes, different operational tradeoffs.
If you are choosing a model today, the decision is less about which one is smartest and more about which one fits your cost structure, latency requirements, and workflow integration needs. That is exactly the efficiency thesis I have been writing about.
Specialization by task, not superiority across tasks
The benchmark split is revealing in another way.
Claude leads on coding. Gemini leads on math and science. GPT leads on general reasoning and professional tasks. This is not a failure of any one model. It is the beginning of a market segmentation.
Different capabilities are starting to find different natural homes. Production engineering teams may gravitate toward Claude. Research and scientific computing may lean Gemini. Enterprise automation and professional workflows may default to GPT.
That looks less like a winner-take-all race and more like an industry maturing into segments — the same pattern we saw with cloud providers, databases, and programming languages before it.
What converging intelligence means for the efficiency stack
This connects directly to the argument I have been building.
If frontier intelligence is converging, then the differentiator moves down the stack. The companies that win will be the ones that deliver equivalent intelligence with better economics — lower cost, lower latency, deeper integration, more reliable operations.
That is why the TurboQuant story matters more than another benchmark chart. And why the ChatJimmy story matters more than another model release. And why Sora's shutdown is so instructive — it shows what happens when you have frontier capability but the delivery economics are broken.
The convergence at the top of the stack is the strongest possible signal that the competition is moving to the middle and bottom of the stack. Compression. Memory architecture. Inference hardware. Deployment tooling. Product experience.
The real winners of March 2026
The most important thing that happened in AI this month was not that GPT-5.4 beat Gemini on some benchmarks and lost on others. It was the pattern underneath: the frontier is getting crowded, the models are getting similar, and the value is migrating from raw capability toward delivery.
The next wave of AI competition will not look like the last one.
The last wave was about who could build the smartest model. The next wave is about who can deliver intelligence — reliably, cheaply, and at the specific shape that actual workflows require. The model becomes a commodity. The stack around the model becomes the moat.
When models converge, what actually wins? The system that makes intelligence disappear into the workflow. Not the one that makes intelligence impressive on a leaderboard.
That is the shift. And March 2026 is when it became undeniable.