The AI coding wars have reached the stage where everybody has a trophy and nobody can explain why there are so many podiums. Claude Opus 4.6, GPT-5.4 and Gemini 3.1 Pro all have impressive results in March 2026. The frustrating part is that those results do not produce a single, clean answer to “which subscription should I keep?”
That is not proof that the models are identical, or that every gap is within a measured margin of error. It means choosing a coding assistant takes more than finding the highest number in a screenshot. Competition works. Unfortunately, it has not abolished homework.
The Numbers Don’t Lie, But the Labels Matter
SWE-bench Verified tests software-issue resolution. Terminal-Bench tests tasks performed through a terminal. ARC-AGI-2 uses abstract reasoning puzzles. They are not interchangeable exams, and an agent’s tools, instructions and allotted effort also affect its score.
Some illustrative results available by March 25:
- SWE-bench Verified: Google’s February model card lists 80.6% for Gemini 3.1 Pro and 80.8% for Opus 4.6. Close numbers on this test are a useful observation, not a declaration that the models are tied at everything.
- SWE-bench Pro, public set: OpenAI reports 57.7% for GPT-5.4. Google reports 54.2% for Gemini 3.1 Pro. These are published results from the respective providers, not a fresh controlled comparison run for this article.
- Terminal-Bench 2.0: OpenAI reports 77.3% for GPT-5.3-Codex and 75.1% for GPT-5.4. Google’s table separately lists results using the Terminus-2 agent setup and other providers’ best reported setups. Mixing those rows into one supposedly uniform race loses an important qualification.
The sources are the Gemini 3.1 Pro model card and OpenAI’s March 5 GPT-5.4 announcement. Think of them as useful evidence with footnotes, not an instruction to buy the winner of every column.
There is a further complication: in February, OpenAI published an audit of SWE-bench Verified identifying test-quality and training-data-contamination concerns. It recommended SWE-bench Pro while acknowledging that Pro is not perfect either. That is a reason to examine the tests, not to replace one sacred scoreboard with another.
And ARC-AGI-2? An interesting reasoning result, certainly. A direct measurement of how well an assistant will fix your authentication bug? No. The trophy cabinet needs separate shelves.
Why There Is No Universal Winner
Changing the agent around a model can change how it searches a repository, edits files, runs tests and recovers from mistakes. Reasoning settings, time limits and available context matter too. A model name alone does not describe the whole working system.
The useful March shortlist includes Opus 4.6 for demanding coding work, GPT-5.4 and GPT-5.3-Codex for coding and agent workflows, and Gemini 3.1 Pro as another serious candidate. That is a shortlist to evaluate, rather than a permanent division in which one company owns architecture, another owns terminals and a third owns reasoning.
A large context window can be helpful, but fitting more files into a prompt does not guarantee understanding every interaction in a repository. The practical question is whether the assistant finds the relevant code and makes a correct, reviewable change. A million-token window is a capacity, not a graduation certificate.
The Price War Is a More Useful Argument
At its February launch, Anthropic listed Opus 4.6 at $5 per million input tokens and $25 per million output tokens. Compare that with the $2 input/$12 output tier for Gemini 3.1 Pro prompts up to 200,000 tokens in Google’s API pricing.
That comparison makes Opus 2.5 times the input price and about 2.08 times the output price. It does not make it ten times as expensive. Nor are input and output interchangeable: a long prompt with a short answer has a different bill from a short prompt followed by pages of generated code.
Those are API token prices, not monthly chat-subscription prices. Caching, long-context tiers, batch discounts, tool use and the amount of generated reasoning can change the bill. The number that matters in practice is closer to cost per accepted result: how much did it cost to obtain a change that passes the relevant checks and is worth keeping?
A cheap model that needs several failed attempts may be expensive. A premium model used to alphabetise a list may be an impressively costly sorting function. Neither outcome needs a brand-war explanation.
Open Weights Change the Options, Not the Electricity Bill
Qwen3-Coder-Next is an open-weight coding model with 80 billion total parameters and 3 billion active parameters. The active count helps describe its computation; it does not mean the full model has the storage requirements of a three-billion-parameter download.
MiniMax’s M2.5 announcement reports 80.2% on SWE-bench Verified and describes its regular and faster Lightning offerings. Again, that is the developer’s published result, with the same benchmark caveats as the closed-model scores above.
Being able to run weights yourself is valuable for control and experimentation. It is not the same as running them for free on any home computer. Hardware, memory, electricity, software setup and maintenance still exist. The GPU does not accept “but the weights were free” as payment.
The Routing Idea: Pick by Task
A multi-model setup sends different jobs to different models, or tries another model when an initial attempt fails. IDC’s routing analysis predicts that by 2028, 70% of top AI-driven enterprises will use advanced multi-tool architectures to manage routing. That is a forecast about a specific group, not evidence that 70% of all businesses already do it.
The logic is simple: you do not use a sledgehammer to hang a picture frame. A sensible experiment might send straightforward drafting to a cheaper model and reserve a stronger configuration for difficult changes. But “cheap”, “mid-tier” and “premium” are choices for a workload, not job titles permanently assigned to particular brands.
A tiny routing demo can be easy to write. A dependable system also needs useful evaluation, sensible fallbacks and a way to notice when the bargain option quietly produces worse work. There is no automatic 60–85% saving with no quality loss, and no reason every individual developer needs to build a router.
What This Means for the AI Labs
Even without a universal capability tie, pricing and the surrounding product matter. Integration depth. Developer tooling. Support. How well the API handles 3 a.m. spikes. Whether the assistant fits the way a team reviews and tests code. The stuff that does not fit neatly into a benchmark bar.
All three companies can find results to emphasise in a launch announcement. That does not make every marketing claim equally broad or every model equally suitable. It makes reading the small print a surprisingly practical developer skill.
What to Try on Your Own Work
If you are choosing a coding assistant in March 2026, start with a few representative jobs rather than an allegiance:
- A bug with a known cause and a test that catches it.
- A feature touching several related files.
- A task involving your actual terminal tools and project conventions.
- A routine job where a cheaper option might be sufficient.
Compare correctness, review effort, time and total cost. Keep the model and agent settings in the record so you know what you actually tested. For more context on the specialist-tool side, see our look at Cursor’s own coding model.
The useful outcome of the AI coding wars is not a permanent winner. It is permission to stop arguing about which model is best and ask which works best for this particular job. Your repository is allowed to have a different opinion from the internet.
