AI・テクノロジー
Gemini 2.5 Ultra Officially Launched — Surpasses GPT-5 and Achieves 2-Million-Token Context Simultaneously
機械翻訳 / Machine-translated
機械翻訳 / Machine-translated
On August 30 at 10:00 PM Japan Time, Google DeepMind officially launched "Gemini 2.5 Ultra." With a context length exceeding 2,097,152 tokens, scores of 92.3% on MATH-500 and 89.7% on HumanEval, the company announced that it "outperforms GPT-5 on 6 out of 9 major benchmarks." The API is priced at $3.50 per 1M input tokens and is available immediately. This will likely be remembered as the day long-context single-model processing moved from "experimental" to "a genuine option."
At 9:47 PM on August 30 (Japan Time), Google DeepMind's official X account announced the release with the single line: "Gemini 2.5 Ultra is here." It became available the same day on Google AI Studio and Vertex AI.
Key specifications are as follows:
"I fed it an entire codebase in a single request for review, and it came back with annotated feedback on a 2,000-line PR in 3 minutes. This is ready for production use."
Multiple engineers posted reports like this on X, and whether the model can be applied to real-world work immediately became the central question.
Since the release of Gemini 2.5 Pro in February 2026, approximately six months of development led to the launch of Ultra. As the LLM competition intensified after OpenAI released GPT-5 in April 2026, Google had published a teaser video for Ultra at "Google I/O Extended 2026" in July, hinting at an August release.
While Anthropic has laid the groundwork for multi-agent distributed processing with its "Claude Agent SDK," Google has doubled down on the approach of "handling long contexts entirely within a single model," making the divergence in architectural philosophies increasingly clear.
If a single A4 page is roughly 700 tokens, then 2 million tokens corresponds to about 2,860 pages. The model has now reached a level where legal documents, large codebases, and entire quarterly reports can be fed in at once. There will be cases where the front-end cost of the conventional "summarize → RAG pipeline → generate" design can be reduced. However, whether the loss of attention in the middle of a long context (the "Lost in the Middle" problem) has been fully resolved is a separate matter, and verification in real-world operations is still necessary.
Gemini has surpassed OpenAI's previously leading math reasoning score (MATH-500: 89.1%). However, OpenAI maintains its o3-series reasoning models as a separate line, and it is important to note that the definition of "best general-purpose model" is now fragmented across different model families. A simple apples-to-apples comparison is already difficult.
The price is slightly above Claude 3.7 Sonnet ($3.00) and well below GPT-5 ($5.00). As price competition heats up at the high-performance tier, Google has staked out a position on both performance and cost. Whether this level is sustainable depends on the inference infrastructure costs of each company.
Existing Google Cloud customers can access the model through Vertex AI without any additional procedures. The model is expected to rapidly penetrate the enterprise market, particularly for large-document processing in healthcare, finance, and legal sectors.
We believe that this announcement has dismantled the prevailing assumption that "long-context models have not yet reached a practical level." Many design patterns written in an era of ~2,000-token cutoffs will likely need to be reconsidered going forward.
At the same time, there are points worth examining with a cool head. Google claims "99.2% on the Needle in a Haystack (NIAH) test," but the distribution of synthetic benchmarks and real-world production data differ. As context length increases, so does inference latency, meaning suitability for real-time tasks must be evaluated separately.
The fact that Google has overturned GPT-5's scores on major benchmarks in just four months since that model's release plainly illustrates a structural shift in the LLM market: model refresh cycles have shortened to six months or less. We will continue to track which players make their next move before the end of the year.
With the launch of Gemini 2.5 Ultra, an era has arrived in which long context, high accuracy, and low cost are achievable within a single model. The competitive advantage over GPT-5 is real on the benchmarks, but practical evaluations covering inference speed, stability, and fine-tuning support are yet to accumulate.
If your system design includes a step for "splitting documents before inputting them," it is worth revisiting that assumption this week.
This article was written by the AI writer (AI News) of the Mirai News editorial team.
