The models did not win by getting bigger. They won time — time to think before answering.
That is the shift that has defined artificial intelligence over the past two years. Reasoning models spend extra computation working through problems step by step: checking their own logic, calling tools, revising their work. By mid-2026, that architecture had produced systems that write production software and finish multi-hour tasks with minimal supervision.
Three labs lead the field — OpenAI, Google DeepMind, and Anthropic — converged on the same core insight, then split sharply on implementation, as researcher jamoeight's May 2026 survey in the open-source "claude-code-deep-research-v2" project documents.
OpenAI's o1, in September 2024, introduced reinforcement learning over hidden chains of thought. Its successor o3 pushed ARC-AGI scores to 75.7–87.5 percent. Then, in January 2025, DeepSeek-R1 showed that pure reinforcement learning could produce reasoning behaviour in a fully open-weights model. That result reset assumptions about how much secret sauce was involved. If an open model could do this, reasoning was not a private trick. It was a technique, and techniques spread.
By early 2026 the family tree had filled out: Claude 3.7 Sonnet's hybrid thinking mode, Gemini 2.0 Flash Thinking, Stanford's s1-32B — 1,000 training samples plus "budget forcing" could elicit reasoning — and Microsoft's rStar-Math, with Monte Carlo tree search at inference time (researcher wanshuiyin's survey).
Three scoreboards, no single winner
The jamoeight snapshot reads like a dead heat, and a strange one — because each lab is winning a different race.
Anthropic's internal Claude Mythos Preview led on software engineering (93.9 percent on SWE-Bench Verified), on general reasoning (64.7 percent on Humanity's Last Exam without tools), and on cybersecurity tasks (100 percent on Cybench). It also showed a demonstrated autonomous working horizon of at least 16 hours, as measured by METR, the nonprofit that evaluates how long AI agents can work independently.
OpenAI's GPT-5.5 and GPT-5.5 Pro led a different set: 90.1 percent on BrowseComp, the deep web research benchmark; 85 percent on ARC-AGI-2, effectively solving it; 82.7 percent on Terminal-Bench 2.0; and 52.4 percent on FrontierMath — alongside a reported 60 percent reduction in hallucinations.
Google DeepMind led somewhere else again: Gemini 3.1 Pro on GPQA Diamond, at 94.3 percent, the graduate-level science benchmark, and Gemini 3 Deep Think — its "parallel thinking" architecture explores multiple solution paths simultaneously — with a gold medal at the 2025 International Mathematical Olympiad, solving 5 of 6 problems.
So who won?
Nobody. And that is the finding.
Read the three scoreboards side by side and a pattern shows. Each lab leads where its architecture leans: Anthropic on software engineering and long runs, OpenAI on web research and terminal work, DeepMind on science and mathematics.
The survey's conclusion goes further: the top three models are now statistically interchangeable on many benchmarks. So harness design — the tools, prompts, and workflows around the model — and cost management outweigh raw model selection for most real deployments.
The model is becoming the commodity. The system around it is the product.
That should change how buyers behave. Picking a winner from one benchmark table is a weak way to decide when the top models are interchangeable on many measures. What surrounds the model — tools, checking, spending caps — becomes the decision that shapes results. It is also the backdrop to the first regulatory probe into rogue AI agents: as systems act longer on their own, the scaffolding around them is where oversight has to live.
One dial, three different machines
The most consequential convergence is invisible to end users: all three labs settled on adjustable reasoning depth as the primary control, each decomposing it differently.
Anthropic shipped Claude Opus 4.7 with a fifth effort tier called "xhigh" and adaptive thinking that replaces fixed token budgets. It also added a Task Budgets API that caps total agentic-loop spend — thinking, tool calls, and output combined — with a model-visible soft limit, enabling self-paced execution.
Pause on that design. The cap is visible to the model itself, so the system can pace its work against a budget it can see instead of burning through computation until an external limit cuts it off — a difference that bites on long tasks, where spending everything early leaves nothing for checking the result.
OpenAI split a different pair of variables: on the GPT-5 family, reasoning effort controls hidden compute while verbosity controls output length independently, and for research agents low verbosity plus high reasoning effort yields the optimal signal-per-token ratio.
DeepMind went a third way. Parallel thinking in Gemini 3 Deep Think explores multiple candidate solutions simultaneously before committing — the same structural idea as the "counterfactual frame ensemble" prompting technique from academic literature, now validated as a first-party architecture.
Three labs. Three decompositions. One shared admission: how a model spends its thinking matters as much as how much of it there is.
It also means "turn the reasoning up" is not a single instruction anymore: up could mean a higher effort tier, a larger task budget, or several paths explored at once.
An agent that writes its own papers
The most striking demonstration of autonomous reasoning in 2026 came not from a coding benchmark but from mathematics.
Google DeepMind's Aletheia — documented in the ai-agent-research daily reports compiled by researcher wbingli — is an autonomous research agent built around an agentic loop. A Generator proposes. A Verifier checks. A Reviser improves. All of it is powered by Gemini Deep Think, with web search to prevent citation hallucinations.
It belongs to the same movement pushing multi-agent systems out of pilots and into production. Except Aletheia has no production line. It has a library.
The results are difficult to overstate, so let the numbers do it. Aletheia scored 95.1 percent on the IMO-Proof Bench Advanced — nearly double the previous record of 65.7 percent — and the January 2026 version of Deep Think uses 100 times less compute than its 2025 predecessor for IMO-level problems.
And beyond benchmarks, Aletheia contributed to peer-reviewed papers, including one fully autonomous paper, designated Feng26, produced with zero human intervention.
DeepMind grades autonomy from Level 0, human-led, to A2, independent and publishable. Aletheia reaches A2: it generates hypotheses, verifies them with online tools, conducts self-review, and produces publication-ready manuscripts — what the lab calls "knowledge generation," a step beyond problem-solving.
The insight at its core travels well. Models reliably catch flaws during dedicated verification that they miss during generation. That pattern is now being ported to coding agents and, increasingly, to the governance frameworks racing to set rules for systems that act on their own. Generate, then check, then revise — as separate steps.
If Aletheia showed what reasoning agents can do given hours, OpenAI's GPT-5.3-Codex-Spark showed what they feel like in real time.
Announced via openai.com and widely discussed (807 points on Hacker News), Spark is a smaller GPT-5.3-Codex variant optimised for Cerebras wafer-scale hardware, delivering over 1,000 tokens per second with a 128K context window.
Traditional coding agents feel like batch jobs: you dispatch a task and wait. At a thousand tokens per second, the agent feels like a collaborator — speed stops being a convenience and becomes a different way of working.
Measuring all this required new instruments, since single-turn benchmarks like MMLU have faded for agentic evaluation. A community-maintained skill guide by Hermes Agent catalogues the battery: GAIA, WebArena, OSWorld, ToolBench (16,000+ instruction-tool pairs), BrowseComp, and Tau-Bench.
And one metric has become the single most cited proxy for real-world autonomy: the METR time horizon — how long an agent can work productively without human intervention. Mythos Preview's 16-plus-hour horizon represents roughly a doubling every year since the metric's introduction, with obvious implications for knowledge work and for how far inference can spread onto decentralised networks running outside the big clouds.
What the contest is really about now
With base capabilities converging, competition has shifted to four axes.
First, efficiency: more reasoning per dollar, where Deep Think's 100× compute reduction and Codex's commodity pricing are opening salvos.
Second, autonomy duration: the METR horizon race, where each additional hour of unsupervised competence unlocks new categories of delegable work — one hour buys a task, sixteen buys a project.
Third, verification architecture. The Generator-Verifier-Reviser pattern from Aletheia is becoming standard, because generation without verification does not scale to high-stakes tasks. A model that produces fluent work it never checks will eventually produce fluent work that is wrong.
Fourth, tool integration. BrowseComp, Terminal-Bench, and OSWorld scores measure how well models use the software environment around them — the difference between a brilliant oracle and a useful employee.
Three caveats keep the picture honest.
Benchmark scores are self-reported by labs, or compiled by third parties from lab disclosures. Independent replication lags, and benchmark gaming remains a live concern. A scoreboard is evidence, but it is evidence largely supplied by the competitors.
Autonomy also does not travel freely between domains. Aletheia's A2 performance is in formal proof; an agent with equivalent reliability in legal negotiation or medical diagnosis has not been demonstrated. Mathematics offers clean verification; a negotiation does not hand you a proof to check.
And the cost of reasoning is real. Every thinking token costs compute, and the most capable configurations are also the most expensive. Read whole, the survey points to a less glamorous challenge for the next two years: making the smartest configurations affordable enough to deploy everywhere.
None of this makes the benchmark tables useless. It changes what they are for. A table can show where an architecture leans and where a system's strengths have actually been measured. What it cannot do, on the survey's own evidence, is hand a buyer a single name and end the argument — the argument now continues into the harness, the budget controls, the verification steps, and the tools on offer.
That is a quieter contest, with no single dramatic scoreboard — but it decides whether reasoning models remain a showcase for the few or become ordinary infrastructure for everyone else.
Reasoning models earn their keep when they stop answering questions and start completing processes. That is the step autonomous multi-agent systems are taking as they move from pilots to production. The management lesson translates from the human version of the same experiment: asynchronous remote enterprises eliminated meeting fatigue by making work legible in writing, and agent systems need exactly that legibility to be supervised at all.
Sources
- jamoeight/claude-code-deep-research-v2, "SOTA Agents 2026 Report" (GitHub, May 2026) — frontier model benchmark snapshot, reasoning-effort comparison: github.com
- jamoeight/claude-code-deep-research-v2, "Data Summary" (GitHub, 2026) — model release/pricing snapshots, GPT-5.5/Gemini 3.1/Claude figures: github.com
- wbingli/ai-agent-research, daily report 2026-02-13 (GitHub) — DeepMind Aletheia details, GPT-5.3-Codex-Spark on Cerebras: github.com
- thecosmicmeta.com, "Google DeepMind Introduces Aletheia" (2026) — Aletheia autonomy levels, FirstProof results: thecosmicmeta.com
- arm2arm/astroagentassistant, "Agentic LLM Evaluation Benchmarks" (GitHub, July 2026) — GAIA, WebArena, OSWorld, ToolBench, BrowseComp catalog: github.com
- wanshuiyin/aris-in-ai-offer, "Reasoning Models Tutorial" (GitHub, 2026) — full survey of reasoning models o1 through rStar-Math: github.com
