Gemini 3.5 Flash review, pricing and verdict

Gemini 3.5 Flash suits agentic coding teams and MCP tool-orchestration pipelines that outgrew Gemini 3 Flash's slower reasoning, trading some abstract-reasoning strength for speed: it outputs at 289 tokens per second, roughly 4x faster than GPT-5.5 and Claude Opus 4.7. Teams chasing the top ARC-AGI-2 score should look at Gemini 3.1 Pro instead.

Google DeepMind's Flash-tier multimodal model spans text, image, audio, video, and PDF inputs in one request, released to general availability in May 2026. It scores 78% on SWE-bench Verified and leads tool orchestration with an 83.6% MCP Atlas score, the highest recorded as of June 2026.

Provider: Google · Family: Gemini 3.5

More about Google on HokAI

Context window: 1,048,576 tokens · Max output: 65,536

Input modalities: text, image, audio, video, pdf · Output: text, tool-calls

About Gemini 3.5 Flash

Gemini 3.5 Flash is Google DeepMind's Flash-tier multimodal model, announced and released to general availability on May 19, 2026 at Google I/O. It succeeds Gemini 3 Flash (released December 17, 2025) and is built on that model's reasoning foundation, adding configurable thinking levels to balance quality, cost, and latency. Google has not disclosed architecture details or parameter counts, consistent with prior Gemini releases; it is a natively multimodal Transformer-based model. Within the Gemini 3.5 family, Flash sits below Gemini 3.5 Pro, announced the same day and still in limited Vertex AI enterprise preview as of June 2026. Beyond its headline coding score, Gemini 3.5 Flash posts 90.4% on GPQA Diamond and 84.2% on MMMU-Pro, the highest multimodal reasoning result Artificial Analysis had recorded at launch. Its Artificial Analysis Intelligence Index of 55 trails GPT-5.5's 60, and on ARC-AGI-2 it scores 72.1% against Gemini 3.1 Pro's 77.1%, showing the hardest abstract-reasoning tasks still favor the Pro tier within Google's own lineup. Claude Opus 4.7 leads the hardest coding benchmark, SWE-bench Pro, at 64.3%. Long-context recall above 100K tokens has not yet been independently published for the 3.5 generation, though Gemini 3.1 Pro demonstrated reliable recall at large depths in prior evaluations, a relevant reference point given the shared Gemini lineage. Gemini 3.5 Flash accepts text, image, audio, video, and PDF input in a single API call but outputs only text and tool-calls; it does not generate images, audio, or video. Function calling and structured output are fully supported, including combined tool use that mixes built-in tools like Google Search grounding and code execution with custom function schemas in the same request, and function-call results can themselves return multimodal objects such as images and PDFs. The model is available through the Gemini API, Google Vertex AI (including a GDPR-compliant EU multi-region endpoint under the Vertex AI Data Processing Addendum), and OpenRouter for third-party API access. There is no open-weights release; Google's separate Gemma family covers that use case. Single-region Vertex endpoints in europe-west3 and europe-west4 still serve only the older Gemini 2.5 and 2.0 models, not 3.5 Flash. Google reports Gemini 3.5 Flash improves on Gemini 3 Flash for both safety and tone while keeping unjustified refusals low, and assesses it unlikely to reach any Critical Capability Levels under its Frontier Safety Framework, based on Gemini 3.1 Pro's evaluation results. Its training data knowledge cutoff is January 2025. Google Cloud holds SOC 2 Type II and ISO 27001 certifications, with HIPAA-eligible configurations available on Vertex AI.

Pricing

$1.50 per 1M input tokens, $9.00 per 1M output tokens. Cached input is $0.15 per 1M tokens, a 90% discount. Roughly 3x Gemini 3 Flash's $0.50/$3.00 pricing but about a third of Claude Opus 4.7's blended cost.

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• 1M-Token Multimodal Context: Accepts up to 1,048,576 tokens of text, image, audio, video, and PDF in one request, with up to 65,536 tokens of output, half of the 2M-token window Gemini 3.5 Pro targets in Vertex AI preview.
• MCP Atlas-Leading Tool Orchestration: Combines native tool calls (Google Search grounding, code execution) with custom function schemas in one request, topping MCP Atlas tool-orchestration rankings as of June 2026.
• Configurable Thinking Levels: Set reasoning depth via a minimal/low/medium/high enum instead of the old integer thinking_budget; 3.5 Flash defaults to medium rather than Gemini 3 Flash Preview's high default.
• 289 Tokens/Second Output: Streams responses at 289 tokens per second, among the fastest output speeds of any model at Flash-tier pricing in June 2026.
• Cached-Input Discount: Cached input tokens are heavily discounted versus standard pricing, cutting repeat-context costs for agent loops that resend the same system prompt or documents.

Pros

• Flash-tier pricing paired with a field-leading MCP Atlas tool-orchestration result achieved in June 2026, an unusual combination of speed and agentic accuracy.
• The highest MMMU-Pro multimodal reasoning score Artificial Analysis has recorded at launch, at roughly a third of Claude Opus 4.7's per-token cost.
• Output throughput is fast enough for high-volume, latency-sensitive agent loops where GPT-5.5 and Claude Opus 4.7 run visibly slower.

Cons

• Output is text and tool-calls only, no native image, audio, or video generation.
• 72.1% on ARC-AGI-2 trails Gemini 3.1 Pro's 77.1%, so Gemini 3.1 Pro remains the stronger choice for pure abstract-reasoning workloads.
• Time-to-first-token stretches to roughly 17.75 seconds at thinking_level='high', compared with well under 5 seconds at medium or low.

Benchmarks

• MMMU-Pro: 84.2% vendor-reported · 19 May 2026 — College-level questions that need reading images and diagrams, % correct.
• ARC-AGI 2: 72.1% vendor-reported · 19 May 2026 — Abstract visual puzzles built to resist memorisation, % solved.
• MCP Atlas: 83.6% vendor-reported · 19 May 2026 — Tasks solved by calling external tools over the Model Context Protocol, % solved.
• GPQA Diamond: 90.4% vendor-reported · 19 May 2026 — PhD-level science questions that are hard to search for, % correct.
• SWE-bench Verified: 78% vendor-reported · 19 May 2026 — Real GitHub issues fixed end to end, % solved.
• AA Intelligence Index: 55 cited: Artificial Analysis · 01 Jun 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
• Output speed: 289 tok/s cited: Artificial Analysis · 01 Jun 2026 — Median tokens written per second as measured by Artificial Analysis.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit Gemini 3.5 Flash Official Page