AI・テクノロジー
OpenAI "o4" Officially Released — 40% Improvement Over o3 in Math & Science Reasoning, Compound Agent Development Enters Full Swing
機械翻訳 / Machine-translated
On September 6, 2026 (local time), OpenAI officially released its reasoning-specialized model "o4" via API and the ChatGPT Pro plan. It posted dramatically higher scores than its predecessor "o3" on compound reasoning tasks spanning mathematics, science, and coding, while API pricing was simultaneously cut by 40%. We have entered a moment where the foundational cost assumptions of agent design are being rewritten once again.
According to OpenAI's official technical report, o4 scored 98.1 points on AIME 2026 (o3: 91.3 points) and 72.4% on SWE-bench Verified (a coding evaluation benchmark) (o3: 63.8%). On "GPQA Diamond," a benchmark for scientific paper analysis, it also achieved 82.7%, representing an 11-point improvement over o3 in accuracy on doctoral-level questions.
API pricing is set at $6/MTok for input and $24/MTok for output. This is a 40% reduction from the previous generation o3 (input: $10/MTok), and the effective cost is expected to decrease even further for tasks with high Chain-of-Thought (CoT) token consumption. The context length remains 200K tokens, and it also supports multimodal input including images, PDFs, and code.
"With o4, I generated 40 compound synthesis routes overnight. That would have been two to three researchers working for two weeks last year. The cost came to just a few dozen dollars." (Pharmaceutical startup researcher, via X)
Since releasing o3 in late 2025, OpenAI has been updating its reasoning model line on an approximately six-month cycle. This release of o4 appears to be a response to the competition in reasoning performance against frontier models such as Claude 5 Sonnet (Anthropic, $3/MTok) and Gemini 2 Ultra (Google DeepMind).
In particular, the SWE-bench score has become the de facto industry-standard metric for autonomous code generation, and with multiple providers now exceeding 70%, the likelihood that a practical baseline for "autonomous code-repair agents" will be established within 2026 has grown significantly. On the pricing front, the input cost for reasoning models overall has dropped roughly 60–70% compared to a year ago, making the cost-effectiveness of multi-step workflows — in which agents make thousands of LLM calls — commercially viable.
GPQA Diamond's 82.7% is just 2.3 points away from the 85% threshold considered "outside-domain expert level." Implementing an "AI researcher" in drug discovery, materials science, and climate modeling domains has become a realistic option.
Exceeding 70% serves as a benchmark for "completing more than 70% of error-correction tasks without human involvement." Combined with the enterprise GA of GitHub Copilot Agent (August 2026), it becomes possible to construct a pipeline that minimizes human intervention from issue filing through PR creation, testing, and merging.
Normally, higher-accuracy models come at higher cost, but o4 achieved both a performance boost and a 40% price reduction simultaneously. While it is more expensive than Claude 5 Sonnet ($3/MTok), the optimal balance of "reasoning accuracy vs. cost" by use case is diversifying — a dynamic that encourages a shift from designs that rely on a single model for all tasks to designs that orchestrate multiple models.
In the technical report, OpenAI referenced an integration schedule with the Codex agent platform. If o4 is adopted as the backend reasoning engine for Codex, it is expected to raise the baseline reasoning accuracy of autonomous software development agents. This can also be read as evidence that OpenAI's strategic axis is shifting from "standalone model" to "vertically integrated model plus agent runtime."
Multiple users have reported unofficial benchmark improvements (Japanese MT-Bench) of 8–12% over o3, suggesting potential gains in practical applicability for use cases such as legal interpretation, medical record processing, and business document summarization. Since no official evaluation has been published, this remains a point requiring continued monitoring.
Updates to reasoning model scores are becoming "scheduled events that arrive every six months." But what deserves attention with o4 is not the benchmark numbers themselves — it is the structural shift that simultaneous improvements in price and performance are becoming the norm.
A year ago, the cost of using high-performance reasoning models carried too much risk for production integration. If that same capability can now be used at $6/MTok with 200K context, the assumptions underlying product design change fundamentally. Particularly in workflows where agents autonomously cycle through hundreds of reasoning steps, cost reductions per step translate directly into the final product price.
The disclosure of the Codex integration roadmap is also not to be overlooked. The fact that OpenAI has begun officially articulating the design philosophy of its entire agent framework — rather than just releasing standalone model notes — can be interpreted as a clear signal of intent to vertically integrate the reasoning engine with the agent runtime. This stands in contrast to the approach taken by Anthropic and Google, which treat APIs and agent platforms as separate layers.
From a developer's perspective, we have now entered an era of selectively using the triangle of o4, Claude 5 Sonnet, and Gemini 2 Ultra depending on the task at hand. Rather than "which is the most powerful," the central axis of design becomes "which model to assign to which task." The decision-making cost of model selection itself will likely become one of the competitive differentiators going forward.
With the release of o4, reasoning models have arrived at a stage where their positioning shifts from "experimental tools for specific use cases" to "the core of production agents." The next focal points are when OpenAI will formally integrate o4 into the Codex and Operator platforms, and whether Anthropic and Google will maintain comparable levels of performance and pricing. The competition over reasoning costs is expected to accelerate further heading into the end of 2026.
Does your organization's model selection criteria for agent design need to be revisited in light of today's pricing levels?
This article was written by an AI writer (AI News) from the Mirai News editorial team.
