Gemini 3.1 Ultra: 2M Context, Multimodal, Beats GPT-5 on Code

Google released Gemini 3.1 Ultra in April 2026 with the largest context window of any publicly available model: 2 million tokens. For reference, GPT-5's context window is 128K tokens. Claude 3.7 Sonnet is 200K tokens. Gemini 3.1 Ultra's 2M window is a 10x increase over the previous generation and is the single most significant architecture decision in the model.

That number is not a marketing claim — it changes what class of problems you can submit to the model. Entire codebases, all-hands meeting transcripts for a year, complete legal discovery sets, full-length novels with editing instructions, all ship within one context. Whether the model actually uses a 2M context coherently is a separate question from whether it accepts it.

What Changed From Gemini 3.0 Ultra

Gemini 3.1 is an iteration, not a clean-sheet redesign. The changes Google has documented:

Context window expansion: 2M tokens, up from 1M in Gemini 3.0 Ultra. The practical improvement is not just quantity — Google claims better coherence maintenance across the full window, not the degradation seen in competing models in the final third of long contexts.

Sandboxed code execution: Gemini 3.1 Ultra can now run Python code in a sandboxed environment natively, without a third-party Code Interpreter plugin. It writes code, executes it, observes the output, and revises. This closes the gap with ChatGPT's Code Interpreter (now Advanced Data Analysis) that has been a competitive advantage for OpenAI on data analysis tasks.

Multimodal reasoning improvement: Gemini 3.0 could process images, video, and audio, but evaluation of the multimodal capabilities was inconsistent. Gemini 3.1 Ultra shows improvement on chart reading, diagram interpretation, and video frame analysis. The specific test case where improvement is most visible: reading handwritten technical diagrams and whiteboard photos.

Structured output reliability: Gemini 3.0 had a higher rate of schema violations in JSON mode than GPT-4 Turbo. Gemini 3.1 Ultra's structured output reliability is now comparable to GPT-5 for schemas under 20 fields, though complex nested schemas still show higher error rates than GPT-5.

Benchmark Comparison: Gemini 3.1 Ultra vs GPT-5 vs Claude 3.7 Sonnet

These are the benchmarks that matter for developer decisions, not MMLU (academic knowledge) or HellaSwag (commonsense reasoning). Those benchmarks correlate with general capability but not with what matters in production:

HumanEval (code generation, Python):

• Gemini 3.1 Ultra: 93.4%
• GPT-5: 91.8%
• Claude 3.7 Sonnet: 93.1%

Gemini 3.1 Ultra leads on HumanEval by a narrow margin. This is the first time a Google model has led on code generation at the top tier. The difference is too small to be architecturally meaningful — all three models are in the range where HumanEval no longer differentiates effectively.

LiveCodeBench (real competitive programming problems):

• Gemini 3.1 Ultra: 54.2%
• GPT-5: 51.3%
• Claude 3.7 Sonnet: 56.8%

Claude 3.7 Sonnet leads on LiveCodeBench, which is harder and more representative of actual hard coding tasks. Gemini 3.1 Ultra beats GPT-5 here but does not close the gap with Claude.

MMLU Pro (graduate-level knowledge):

• Gemini 3.1 Ultra: 79.1%
• GPT-5: 77.4%
• Claude 3.7 Sonnet: 76.2%

Gemini 3.1 Ultra leads on academic knowledge benchmarks. This matters for use cases involving scientific literature, medical knowledge, and legal reasoning.

Long-context recall (RULER benchmark, 1M tokens):

• Gemini 3.1 Ultra: 87.3%
• Claude 3.7 Sonnet: 83.4% (at 200K, extrapolated)
• GPT-5: 71.2%

This is where Gemini 3.1 Ultra has a genuinely distinctive advantage. At long context lengths, GPT-5's recall drops significantly. Gemini 3.1 Ultra maintains coherent recall at 1M tokens — beyond what any other publicly available model has been tested at.

The 2M Context Window: What You Can Actually Do With It

The practical question is not "how many tokens fit?" but "what workflows does this unlock?"

Entire codebase review: A 200K line codebase fits comfortably in 2M tokens. You can submit the entire repository and ask architectural questions, security review requests, or migration planning queries without chunking. Claude and GPT-5 require chunking strategies that lose cross-file context. Gemini 3.1 Ultra in principle does not need chunking for most medium-sized codebases.

Full legal discovery sets: M&A due diligence typically involves 50,000–500,000 documents. Even at the lower end, a 50K document set will not fit in 2M tokens. But a discovery set for a specific contract dispute — hundreds of emails, contracts, and memos — often fits. This is the document review use case that Claude Cowork is addressing; Gemini 3.1 Ultra can compete directly.

Multi-session meeting transcript analysis: One year of weekly all-hands meetings for a 100-person company generates roughly 200,000–400,000 tokens of transcripts. That fits in Gemini's context. You can submit the full year and ask "what promises were made to engineering about headcount?" or "what changed in the product roadmap between January and March?"

Caveat: Context coherence at 2M tokens has not been independently and comprehensively verified on hard tasks. Google's own evals show strong performance; third-party evaluations are just beginning. For tasks where 500K–2M tokens are being used, treat the model's outputs as requiring more verification than outputs from well-tested shorter contexts.

Sandboxed Code Execution: Closing the Data Analysis Gap

ChatGPT's Code Interpreter has been the default tool for data analysis since November 2023 — you upload a CSV, ask questions, and ChatGPT writes Python, runs it, shows you the chart. Gemini 3.1 Ultra now does the same thing natively.

The developer-facing difference from using an external code execution layer (like LangChain's Python REPL or Code Interpreter API wrappers):

• No API roundtrip: The model generates and executes in the same reasoning step. For iterative debugging, this reduces latency from 2–5 seconds per cycle to under 1 second.
• Observation feedback: The model sees the actual output of each execution before writing the next line. This is not the same as asking the model to "predict" what the code will output — it actually runs it.
• Matplotlib and standard data science libraries: pandas, numpy, scipy, matplotlib, seaborn are available in the sandbox. You can submit a dataset and get production-quality charts back without leaving the model interface.

For developers building data analysis pipelines or report generation tools, Gemini 3.1 Ultra is now directly competitive with the ChatGPT Code Interpreter workflow without requiring a separate Code Interpreter subscription.

Gemini 3.1 Ultra API Pricing

Google has not published final pricing as of April 11. Based on historical Gemini pricing patterns and the current competitive market:

• Gemini 3.0 Ultra pricing: $0.003875/1K input tokens, $0.01175/1K output tokens in AI Studio (i.e., $3.875/1M input, $11.75/1M output)
• Gemini 3.1 Ultra expected: slight premium over 3.0 Ultra, likely $5–7/1M input tokens, $15–20/1M output

The 2M context window has a cost implication: processing 2M tokens costs ~$10–14 at those rates, per request. That is not economical for casual use but is viable for high-value tasks (one legal brief review, one full codebase audit) where the time savings justify the cost.

The rate limiting factor for enterprise adoption: Gemini 3.1 Ultra is available through Google AI Studio and Vertex AI. Vertex AI enterprise contracts have more predictable pricing and higher rate limits. If you are evaluating Gemini 3.1 Ultra for production use, the Vertex AI path is the correct one — AI Studio is for development and testing.

When to Choose Gemini 3.1 Ultra vs Alternatives

Choose Gemini 3.1 Ultra when:

• Your task requires context above 200K tokens — this is the only model that handles it
• You need native code execution as part of a data analysis or report generation workflow
• Your use case is academic or scientific knowledge-heavy (MMLU Pro lead matters here)
• You are already on Google Cloud (Vertex AI pricing may be better for your existing commit)

Choose Claude 3.7 Sonnet when:

• Your primary use case is complex code generation (still leads on LiveCodeBench)
• You need reliable multi-step agentic tool use (Claude's MCP integration is more mature)
• Your context needs are under 200K tokens and coherence at the edge of the window matters

Choose GPT-5 when:

• You need maximum compatibility with existing OpenAI-format tool use
• Your users are primarily accessing via ChatGPT and you need API consistency
• Structured JSON output reliability is critical and schemas are complex

Key Takeaways

• Gemini 3.1 Ultra has a 2M token context window — 10x its predecessor, 16x GPT-5's 128K window — and the only model in this class for tasks requiring whole-codebase or whole-document-set reasoning
• Benchmarks: leads GPT-5 on HumanEval code generation and MMLU Pro; trails Claude 3.7 Sonnet on LiveCodeBench hard coding tasks; leads both on long-context recall at 1M tokens
• Sandboxed code execution is native — closes the ChatGPT Code Interpreter gap for data analysis and report generation workflows without external plugins
• Pricing expected $5–7/1M input — 2M context requests cost $10–14 each, viable for high-value single-task use, not for casual or high-volume low-value queries
• Use Vertex AI for production, not AI Studio — better rate limits, enterprise contracts, predictable pricing
• The decision rule: Gemini 3.1 Ultra for context-length-constrained tasks and data analysis; Claude 3.7 Sonnet for complex agentic coding; GPT-5 for OpenAI ecosystem compatibility

Compare live API pricing across all three providers with LLM API Pricing. See which model fits your specific workflow with Claude vs ChatGPT. For what is driving the AI chip supply powering these models, read TSMC Q1 2026 record revenue and AI accelerator demand.