Google’s Gemini 3.8 Flash Scores 59 Behind Fable 5.1 and GPT-5.6 Sol

Google introduced Gemini 3.8 Flash and the access-restricted Gemini 3.8 Flash Cyber on September 2, with the general model’s independently measured score of 59 sitting behind Claude Fable 5.1 and GPT-5.6 Sol. At $0.58 per measured task, Gemini cost less than both higher-scoring systems. Developers choosing a model for agent work must weigh lower task cost and faster generation against lower benchmark performance.

Two points behind Sol

On the Artificial Analysis Intelligence Index, accessed September 2, Gemini 3.8 Flash at high effort scored 59 and cost $0.58 per task. GPT-5.6 Sol at maximum effort scored 61 and cost $0.95 per task, while Claude Fable 5.1 at maximum adaptive effort scored 66 and cost $3.69. Fable’s default fallback means its result is not a pure test of one model.

The task prices are weighted averages across the index, not provider list prices. Full-run totals also depend on each system’s token use, so models with identical per-token rates can produce different bills on the same evaluation.

Gemini generated 304.6 output tokens a second and produced 120 million output tokens across the evaluation; the full run cost $825.83. Sol generated 72.4 tokens a second, produced 70 million output tokens and cost $2,017.29. Fable generated 66.4 tokens a second, and its evaluation cost $8,523.16.

In the index snapshot accessed September 2, open-weights GLM-5.3 Flash was cheaper and scored lower. It scored 57, two points below Gemini, but cost $0.09 per task and $138.02 for the full run. Gemini was much faster than GLM’s 42.8 output tokens a second and used fewer than GLM’s 150 million output tokens, yet Gemini’s $0.58 measured task cost was about 6.4 times GLM’s $0.09 cost.

Unchanged rates, higher task cost

In the index snapshot accessed September 2, Gemini 3.8 improved on Gemini 3.7 Flash’s high-effort score of 56. The gain came with a rise in cost per task from $0.40 to $0.58 and an increase in output volume from 64 million to 120 million tokens across the respective evaluations.

The API rates did not change between Gemini 3.7 Flash and Gemini 3.8 Flash: $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026. Those rates double on January 1, 2027. Stable token prices therefore did not produce a stable task cost in the outside test. Actual results can also differ at the model’s default medium effort, since this comparison used high effort.

Google’s model card states the tradeoff directly: “At times, the model might use more tokens to maximize performance, especially at higher effort levels.” Developers can lower the effort setting or continue using 3.7 for work where compute efficiency matters more.

In Finance Agent v2, accessed September 2, 3.8 led three of nine categories, but 3.7 still led general quantitative analysis. On the held-out Harvey legal test that day, 3.8 resolved 10.0% of complete tasks. Muse Spark 1.2 led complete-task resolution at 25.42%.

A separate cyber release

On CWE-Bench, accessed September 2, Gemini 3.8 Flash Cyber passed 47.2% of 100 held-out audit-and-patch tasks, just below Claude Fable 5 at 47.8% (the general Intelligence Index uses Claude Fable 5.1) and above GPT-5.6 Sol at 44.2% and Gemini 3.7 Flash at 44.0%. The deterministic verifier accepts a patch only when the exploit stops working and existing tests still pass.

Google’s own tests put the cyber model above 70% success across 20 programming languages and found 2.6 times as many correct Chrome patches as unnamed larger commercial models. Google did not disclose the underlying test sets or the competitor identities.

The cyber model is not available through the general Gemini API. Fairwind access is limited to approved governments, infrastructure operators, software maintainers and research teams, with named-user controls. Partners may use it for permitted defensive or academic research and cannot share, resell or redistribute access.

As of September 2, CWE-Bench publishes average rollout costs of $10.27 for Fable 5, $2.29 for Sol and $1.43 for Gemini 3.7, but no cost for 3.8 Cyber. Google’s claim that the new model costs “significantly” less therefore has no external figure to test.

Related stories

Google Cuts Gemini 3.7 Flash API Price by 50%

Google released Gemini 3.7 Flash three weeks after its previous model, with API rates at $0.75 per million input tokens and $3.75 per million output tokens through December. Coding and agent benchmarks improved, but Google's own table still gives GPT-5.6 Terra several workload wins.

Implicator.ai

Grok 4.6 Scores 61 at One-Fifth of GPT-5.6 Sol Output Price

Independent testing puts Grok 4.6 level with GPT-5.6 Sol at one-fifth of its standard output-token price. A 32-second latency result and doubled long-context rates set the limits.

Implicator.ai

OpenAI Ships GPT-5.6-Cyber to Vetted Defenders in Daybreak R

OpenAI has released a model trained to answer the cyber requests its consumer model refuses, days after pausing work on Astra. GPT-5.6-Cyber completed 95% of advanced cybersecurity requests in OpenAI's own tests. Every figure is the company's own measurement.

Implicator.ai

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai