Google Unleashes Gemini 3.1 Ultra With a 2-Million-Token Brain That Can Watch Movies Whole

Developers also get a new code-execution tool. Inside a single conversation, the model can write a program, run it on a sandboxed server, see what happens, and then fix any errors. Google says these improvements should reduce 'hallucinations,' the false but confident-sounding answers that AI models sometimes produce.

Level 3 — Intermediate

Google's Gemini 3.1 Ultra, unveiled this week, is the company's biggest model release of 2026 and a clear signal that the multimodal era of artificial intelligence has decisively arrived. The headline specification is a context window of two million tokens — large enough to ingest a feature-length film, roughly fifteen hours of speech, or three full novels in a single prompt without any of the chunking, summarization, or vector-search workarounds that older systems required.

Crucially, Gemini 3.1 Ultra handles text, images, audio, and video natively, without first transcribing speech to written form. That single design choice eliminates an entire class of brittleness: cues such as tone of voice, background noise, and timing — which transcription strips away — are preserved end-to-end. The model can therefore reason about a podcast monologue the same way a human listener might, hearing sarcasm, pauses, and changes of speaker.

Developers receive a substantial upgrade in agentic functionality. A new sandboxed Code Execution tool allows Gemini 3.1 Ultra to author Python or JavaScript, run it on a locked-down server, inspect outputs and stack traces, and self-correct mid-conversation. Combined with improved grounding via Google Search, this places the model in a small group capable of conducting genuine research workflows rather than producing static answers.

Industry analysts say the launch repositions Google sharply against OpenAI's GPT-6 and Anthropic's Claude Opus 5, both of which retain leadership on certain reasoning benchmarks but trail Gemini on long-context comprehension. Pricing for the API tier is still below OpenAI's flagship, and Google has bundled the new model into Workspace Enterprise at no extra charge, betting that distribution will matter as much as raw capability.

ingest: To take in or absorb information for processing.

chunking: Breaking long data into smaller pieces so a model can handle it.

brittleness: A tendency to fail abruptly when conditions change.

agentic: Able to plan and take actions independently to achieve a goal.

stack trace: A report showing where in a program an error occurred.

grounding: Connecting an AI's output to verifiable real-world information.

benchmark: A standard test used to compare performance.

distribution: Here, the channels through which a product reaches customers.

Level 4 — Advanced

Google's release of Gemini 3.1 Ultra this week represents the most ambitious frontier-model launch from Mountain View in more than a year, and a pointed corrective to the perception that the company has been ceding multimodal leadership to OpenAI and Anthropic. The headline figure — a native two-million-token context window — is not merely a quantitative bump. It crosses a qualitative threshold at which most chunking, retrieval-augmented generation, and hierarchical summarization scaffolding that has dominated production AI pipelines for two years becomes optional rather than obligatory.

Architecturally, the model is reported to dispense with the audio-to-text transcription pre-step that has hobbled every previous purportedly multimodal system. By preserving acoustic, prosodic, and visual cues end-to-end, Gemini 3.1 Ultra reasons about a podcast monologue or a news broadcast in something closer to the way a perceptive listener would, retaining sarcasm, hesitation, ambient sound, and timing. Early benchmark results on the LongVideoBench and MMAU audio-reasoning suites place it materially ahead of competitors that depend on Whisper-class transcription layers.

For developers, the release significantly extends Gemini's agentic surface area. A new sandboxed Code Execution tool lets the model author Python or JavaScript, dispatch it to an ephemeral Cloud Run container, examine standard-out streams and stack traces, then self-correct — all within a single user turn. Coupled with substantially improved grounding via Google Search and a new structured-citation mode that flags every factual claim with a hyperlink, this elevates Gemini from a static-answer engine into something more closely resembling a junior analyst. Hallucination rates on internally curated factual-recall benchmarks are reported to have fallen by roughly forty percent versus Gemini 3 Pro.

Strategically, the launch sharpens Google's pincer movement against OpenAI's GPT-6 and Anthropic's Claude Opus 5. Both rivals retain narrow leads on certain abstract-reasoning benchmarks, particularly ARC-AGI-2 and the FrontierMath frontier slice, but neither matches Gemini's long-context comprehension or its bundled distribution through Workspace Enterprise — included at no additional cost to existing seat-holders. With API tariffs priced below GPT-6's flagship rate and a new on-device variant earmarked for the Pixel 11 launch in the autumn, Google is wagering that capability plus ubiquity, rather than capability alone, will define the next phase of the AI race.

frontier model: A leading-edge AI model that defines the state of the art.

retrieval-augmented generation: A technique where a model fetches outside documents to improve answers.

prosodic: Relating to the rhythm, stress, and intonation of speech.

ephemeral: Lasting only a very short time.

container: An isolated software environment for running an application.

structured-citation: A format that explicitly links each factual claim to a source.

pincer movement: A strategy of attacking from two directions at once.

ubiquity: The state of being everywhere or used by everyone.

Level 1 — Absolute Beginner

Level 2 — Elementary

Level 3 — Intermediate

Level 4 — Advanced

Google Unleashes Gemini 3.1 Ultra With a 2-Million-Token Brain That Can Watch Movies Whole

Multiple Choice

True or False

Fill in the Blank