Google's Gemini 3.5 Pro Sets 2M-Token Context Standard at $1.25 Per Million Tokens

Google forces enterprise TCO recalculation with concrete long-context pricing

Google published preliminary pricing for Gemini 3.5 Pro at $1.25 per million input tokens and $10 per million output tokens, with a 2-million-token context window now in enterprise preview on Vertex AI. The $250-per-month Ultra tier includes access to a dedicated "Deep Think" reasoning mode. For enterprise buyers budgeting agentic workflows, code generation at scale, or multi-document analysis, this is the first time a hyperscaler has put concrete economics around context windows materially larger than the 1-million-token range that became standard in 2025.

The immediate impact: CIOs can now build actual TCO models comparing Gemini 3.5 against existing OpenAI GPT-4/5 or Anthropic Claude contracts for specific workloads instead of navigating "contact sales" pricing. A 2-million-token context reduces the infrastructure complexity of retrieval-augmented generation (RAG) systems—fewer chunking layers, less orchestration code, lower operational risk around vector databases and embedding pipelines. For workflows processing large codebases, financial filings, or legal discovery sets, the architectural simplification is material enough to justify re-scoping existing LLM deployments.

The tradeoff: Gemini 3.5 Pro remains in enterprise preview, meaning final benchmarks, GA timelines, and locked pricing are not yet committed. Buyers treating this as an immediate procurement standard are taking on roadmap risk. The more rational near-term use is proof-of-concept acceleration in domains where long-context and reasoning depth matter—financial research, complex legal drafting, multi-step coding agents. Organizations that have delayed LLM pilots due to RAG complexity now have a simpler technical path, even if the contractual certainty lags by a quarter.

On benchmarks, Gemini 2.5 Pro with Deep Think leads GPQA Diamond and MMLU-Pro among public cloud endpoints. Google positions 3.5 Pro as the successor in that performance tier, directly competing with OpenAI GPT-5 Pro and Anthropic Claude 5 Opus for high-end reasoning workloads. The 2-million-token window puts pressure on both competitors to match long-document capabilities without forcing enterprises into heavier prompt-engineering overhead.

OpenAI closes reliability gap in production coding and document workflows

OpenAI updated GPT-5 Pro with improved function-calling reliability and multi-turn context handling, targeting issues enterprises reported in complex coding assistance and long-document analysis. The update adds expanded structured output support, making responses more predictable for production pipelines that depend on deterministic JSON or formatted data.

This is a risk-management release, not just a capability upgrade. For enterprises standardizing on OpenAI APIs in finance, healthcare, or regulated sectors, the improved function-calling reliability directly reduces integration risk. Fewer edge-case failures mean less manual guardrail code and fewer human-review checkpoints in LLM-driven apps. Benchmarks cited in coverage show gains on complex coding tasks and long-document analysis—two areas where prior GPT-5 versions lagged under production load.

The competitive context: this update narrows the gap with Anthropic Claude 5 in structured tool use and hallucination control, and with Gemini 2.5/3.5 in long-document and agentic workflows. For buyers already in procurement cycles for GPT-4/5-based platforms, the reliability improvements shift the risk calculus. GPT-5 Pro can now legitimately be treated as production-ready for complex multi-turn coding and document workflows, potentially accelerating final sign-offs where stability concerns had slowed deals.

No new list pricing is reported for this update, but the reliability improvements strengthen the case for continuing or expanding OpenAI contracts rather than shifting to competitors purely for stability reasons. The cost of vendor migration now includes the opportunity cost of losing these production-hardening improvements.

Anthropic reduces hallucinations in legal and medical domains, adds 60-database Science Workbench

Anthropic released a Claude 5 refinement update focused on reducing hallucinations in specialized domains—specifically legal and medical workflows—alongside improved tool-use capabilities and better structured output reliability at scale. Separately, Anthropic launched Claude Science Workbench, connecting Claude Opus 4.8 to more than 60 scientific databases across genomics, proteomics, cheminformatics, and clinical trial literature. Anthropic is backing the Workbench with AI for Science grants covering 50 projects at $30,000 in credits each, roughly $1.5 million in total credits.

For legal and medical enterprises, the hallucination reductions and structured tool use address the primary obstacle to LLM deployment: the need for extensive custom guardrails and human review. This can shift RFP scoring where "model safety and reliability" is weighted heavily and potentially justify premium pricing for Claude over more generic models. The domain-reliability focus competes directly with GPT-5 Pro and Gemini 2.5/3.5 in regulated industries where hallucinations and tool reliability are gating issues for adoption.

The Science Workbench's 60-database connectivity gives research-heavy enterprises a clearer path to centralizing LLM-driven literature review, hypothesis generation, and experimental design without building custom integration layers. This competes with AWS's and Microsoft's emerging AI-for-science platforms, but with a strong emphasis on integrated databases and Opus-level reasoning. For pharmaceutical, biotech, and academic research IT buyers, the Workbench reduces the infrastructure lift of connecting proprietary LLMs to external scientific data—a common blocker in research AI pilots.

What to watch

Google's Gemini 3.5 Pro GA timeline and final pricing lock will determine whether the 2-million-token context becomes a de facto standard or a preview feature that competitors ignore. If Google ships GA in Q3 with pricing near the preview range, expect OpenAI and Anthropic to respond with either matching context windows or sharper pricing on existing ranges to defend share.

OpenAI's GPT-5 Pro reliability update sets a floor for production-grade LLM behavior in enterprise coding and document workflows. Buyers should model multi-turn coding and long-document use cases against this new baseline rather than earlier GPT-5 behavior when evaluating competing models.

Anthropic's domain-specific reliability focus in legal and medical—and the Science Workbench's 60-database integration—signals where vertical LLM competition is heading: not just general reasoning, but pre-integrated tooling for regulated and research-heavy industries. Buyers in those sectors should model the cost of custom integration layers against the premium Anthropic is likely to charge for turnkey domain stacks.