Gemini 3.7 Flash: Real Pricing, Speed, and Agent Boundaries (eesel)

One-sentence takeaway

Gemini 3.7 Flash is very strong on speed, documents/multimodality, and narrow-scope refactoring, but higher thinking levels significantly increase time to first token and token costs; long-term customer-support/automation selection must be dry-run against your own historical tickets.

Use cases

• Good for: PDF/OCR/chart and long-context retrieval; image, video, and document understanding; web/UI generation; narrow-scope code refactoring; high-throughput batch tasks.

• Not good for: Directly replying to customers without human review; pursuing one-shot completion with long-running terminal agents; classification/labeling work where “lowest token price” is the only objective (minimal has been removed and thinking tokens are billable).

• Applicable model version: gemini-3.7-flash, GA release (2026-08-13).

• Applicable clients, agents, or APIs: Gemini API, Google AI Studio, Google Antigravity; the article also compares OpenRouter/Vertex runtime metrics and API pricing.

• Recommended reasoning level and parameters: low is suitable for low-latency drafts, real-time chat, and quick analysis; medium (default) is suitable for complex code/agents; high is suitable for hard math, the most difficult coding, and long tool chains, but requires tolerating higher latency and token consumption.

Test environment

• Independent run data: Artificial Analysis speed/intelligence/hallucination data; OpenRouter P50 and provider throughput; the author's real customer-support integration experience at eesel (more than 100,000 German Zendesk tickets per month for a single customer).

• Publisher comparison: Google's 21-row model-card comparison table, including Gemini 3.6 Flash, Claude Sonnet 5, GPT-5.6 Terra, and Muse Spark 1.2.

• Pricing dates: The introductory price through 2026-12-31 is $0.75/1M input and $3.75/1M output; from 2027-01-01 it becomes $1.50/1M and $7.50/1M. Thinking tokens are billed at the output rate.

Inputs/configuration

• Input modalities: text, images, video, audio, and PDF; text output.

• Context window: 1M tokens; output limit: 64K.

• When migrating from older Gemini versions, the article advises removing temperature, top_p, top_k, and candidate_count, changing numeric thinking_budget to thinking_level, removing prefilled model turns, and moving multi-turn history to the server-side previous_interaction_id.

• The article notes that the minimal thinking level is no longer available in 3.7 Flash; callers should record low/medium/high explicitly and not assume that the old model's four-level configuration remains compatible.

Results data

Speed, cost, and thinking levels

• Artificial Analysis: 340.1 output tokens/s, ranked first in speed among 188 models; the median of the comparison category on the same page was 68.6 tok/s.

• OpenRouter: approximately 93 tok/s P50 and 1.85 seconds latency; Google AI Studio approximately 124 tok/s/1.74 seconds, and Vertex approximately 84 tok/s/2.02 seconds.

• High thinking uses approximately 40% more tokens than Gemini 3.6 Flash, averaging about 37K output tokens per task; being fast does not mean lower total task time or cost.

• Artificial Analysis recorded an independent hallucination rate of 64.5%, versus 55.6% for Gemini 3.6 Flash; the author emphasizes that improved accuracy and a higher hallucination rate can occur at the same time.

Key rows from Google's published table

Conclusions

1. Capability shape is more useful than the overall score: 3.7 Flash shows a consistent advantage on documents, video, charts, long context, and UI generation; all four key indicators for long-horizon terminal coding are still led by GPT-5.6 Terra.

2. Low may be the production default: In the data cited by the author, low scores 50.9, close to 3.6 Flash at its high level, while time to first token is 0.74 seconds and task cost is $0.16; the default medium should not be used unconditionally.

3. Speed must be calculated per task: 340 tok/s output speed is offset by the 37K average output tokens from high-level thinking; track completed-task cost and human corrections rather than looking only at tok/s.

4. Reliability depends on external controls: The author recommends that supported scenarios use a “classify/retrieve → attempt automation → route low-confidence cases to a human” workflow, then dry-run it on your own historical tickets before launch.

Limitations

• Some numbers in the article come from Artificial Analysis, OpenRouter, and Google's published table; they are not a complete independent comparison under the same harness.

• The definition, sample, task set, and time window for the hallucination rate are not fully disclosed in this article; it cannot be directly extrapolated to Tabbit content or customer-support data.

• eesel's production workload and commercial product experience may involve selection bias, and the 100,000+ ticket case is not a publicly reproducible dataset.

• Some items in Google's table, such as Terminal-Bench 3.0 and AutomationBench, use private sets or the publisher's own methodology; do not conflate them with Artificial Analysis's independent measurements.

• The prices are introductory rates through the specified dates; the official pricing page must be checked again for a launch budget.

Reproduction steps

1. In the Gemini API or AI Studio, fix the same input set, explicitly switch among low, medium, and high, and save usage.total_thought_tokens, time to first token, total latency, and output tokens.

2. For the same set of repository refactoring, PDF/chart understanding, video summarization, and JSON structured-output tasks, record success rate, tool errors, human corrections, and total cost per completed task separately.

3. For each API endpoint, record P50/P95 time to first token and throughput for Google AI Studio, Vertex, or OpenRouter; do not infer one provider's speed from another provider.

4. Before launch, dry-run at least 500 closed customer-support tickets: classify answers into can be sent automatically, needs a human draft, and should be escalated, then check retrieval hits, factual errors, and hallucination rate.

5. Recalculate the cost of each task under the two pricing tiers on 2026-12-31 and 2027-01-01, including thinking tokens, cached reads, batch processing, or priority mode.

Original evidence and data

• The eesel page provides a table of the Intelligence Index, time to first token, and per-task cost for the low/medium/high levels, and notes that minimal is no longer available.

• The page reports Artificial Analysis's 340.1 tok/s, 64.5% hallucination rate, and the 55.6% comparison for Gemini 3.6 Flash; these are third-party records, not a benchmark run the author claims to have built.

• The page reproduces Google's 21-row comparison table and explicitly states that 3.7 is slightly below 3.6 on both CharXiv sub-items and trails GPT-5.6 Terra on Terminal-Bench/DeepSWE/OSWorld.

• The page recommends a long-term customer-support dry run using your own closed tickets for validation, rather than going live directly based on benchmarks or a vendor demo.

Source excerpts or observations (compliance short quotes)

“Tokens per second and time to answer, these are not the same thing.”

“Triage, then attempt, then escalate.”