AI・テクノロジー
OpenAI "o4" Officially Launched — Code Reasoning Exceeds 71%, Reshaping the Stakes of Agent Design
機械翻訳 / Machine-translated
On August 16, 2026 (US time), OpenAI simultaneously released its reasoning-focused model "o4" via API and ChatGPT Pro. It achieved 71.3% on the code benchmark "SWE-bench Verified" — a leap of more than 18 points from the previous generation o3's 53.1%. The accuracy rate on math olympiad-level problems (AIME 2026) reached 92.0%, and as of today, the intuitive threshold for "what can be entrusted to an AI agent" has shifted dramatically.
Key benchmarks for o4 announced by OpenAI officially:
The API offers two models: "o4" and "o4-mini." o4-mini is priced at $2.00 per 1M input tokens, and o4 at $15.00. Compared to o3-mini, o4-mini achieves approximately 41% cost reduction.
Immediately after the release, reactions from engineers flooded X.
"SWE-bench 71.3% on o4 is 2–3 quarters ahead of what I expected. The math now works out for autonomous PR merging to become a reality before the end of this year."
— AI engineer account (tens of thousands of followers)
Since the release of o3 (December 2025), the "reasoning wall" had continued to surface as a challenge in real-world autonomous agent implementations. An SWE-bench score in the 53% range effectively meant "human review required for about 80% of tasks," creating a clear gap for fully autonomous code PR generation.
In the meantime, the industry focused its competition on context length, multimodality, and output speed — but o4 stands apart in that it was deliberately optimized around "reasoning depth." Internally, the reinforcement learning-based self-correction loop (Self-Correction) is believed to have been significantly enhanced. From o3 to o4 took roughly eight months — the compression of the competitive cycle has once again been made visible in the numbers.
A SWE-bench score of 71.3% represents a level at which many small-to-medium bug-fix tasks with tests are beginning to fall within the range of "unsupervised execution." In agent workflow design, we have entered a phase that demands a redefinition of the "trigger conditions for human review."
At $2.00/1M tokens, the cost of running reasoning-heavy tasks continuously in the startup and SMB space becomes genuinely viable. The incentive to migrate from o3-mini is substantial, and the composition of API usage is expected to shift over a short period.
As of the same day, SWE-bench comparisons showed Gemini 2.5 Ultra at 67.8% and Claude Opus 4.7 (released July 2026) at 69.2%, with o4 taking the numerical top spot at 71.3%. However, the optimal choice differs depending on the balance of inference latency, context length, and cost structure. Reading "o4 as the optimal solution for all use cases" would be premature.
Domestic system integrators and contract development firms had already been incorporating o3 into code review assistance workflows. Migration to o4 is expected to further increase the reduction rate in review processes, but the majority of companies maintain the stance of keeping security audits and license verification in human hands.
How one reads the figure of 71.3% is the fork in the road. The remaining 28.7% is not "tasks AI is bad at" — it is "problems that cannot yet be solved probabilistically with certainty," and that probability fluctuates from session to session. The appropriate design approach at this moment is a structure that "processes in bulk the tasks with high success probability," not one that "operates in full autopilot mode."
More important than that, however, is the story of the cycle. More than 18 points of improvement from o3 to o4 in approximately eight months. If this pace is sustained, a projection of SWE-bench in the 80% range by the first half of 2027 becomes plausible.
The asymmetry in decision-making costs for enterprises is changing right now. The timing at which the cost of "waiting and watching" begins to exceed the cost of "adoption" is traced by the curve of model performance. Today's announcement will be recorded as a clear inflection point on that curve.
The release of o4 has produced numbers that can serve as the basis for moving autonomous agents from "PoC" to "production workflow." However, the evaluation framework must go beyond accuracy alone — cost, latency, and security requirements form the three axes to consider. The next point to watch is the order in which OpenAI integrates o4 into agent frameworks such as Operator — that will effectively determine the speed at which it permeates practical use.
This article was written by an AI writer (AI News) from the Mirai News editorial team.
