Gemini 3.8 Flash scores 59 on the Artificial Analysis Intelligence Index (high reasoning), up 3 points from 3.7 Flash’s 56, and matches GPT-5.6 Sol (xhigh) and Grok 4.6 (medium).
Introductory API pricing is $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026, rising to $1.50 and $7.50 on January 1, 2027.
Cost per task increased about 40% to $0.58 from $0.40 despite unchanged token pricing, because the model produces about 30% more output tokens per task (48k) and runs more agentic turns.
On DeepSWE v1.1 the model reaches 73.7%, just below Claude Opus 5’s 74.0% and ahead of GPT-5.6 Sol’s 72.7%.
Gemini 3.8 Flash Cyber is restricted to Google’s Fairwind Program and is not publicly accessible.
What Actually Changed vs. 3.7 Flash
Google presents the release focusing on a single design choice: the model performs more work. On complex tasks it carries out additional reasoning steps and calls tools repeatedly, and it may use more tokens to improve performance at higher effort levels. This explains the improvements and also accounts for the unexpected cost increase described below.
What Actually Changed vs. 3.7 Flash, architecture diagram
The context window remains at 1 million tokens, and multimodality still supports text, image, video, and speech input with text output. The change is in agentic capability. Artificial Analysis attributes the three-point index increase mainly to better performance on tool use, coding, and real-world task evaluations, with the largest single gain on the τ³-Banking tool-use benchmark, where the model improved 12 points over 3.7 Flash to reach 45%.
Google’s materials highlight long-horizon software engineering and specialized professional domains. The company reports 54.9% on HLE-Verified, its measure of multi-step reasoning across STEM, humanities, and professional fields, and notes improvements on the Vals Finance Agent V2 and Harvey’s Legal Agent Benchmark suites. These figures come from the launch post; the independent confirmation so far is the Artificial Analysis index, which supports the direction and approximate size of the improvement.
The most important analytical point in this release is the difference between token price and task cost. The per-token price stayed the same: $0.75 input and $3.75 output per million tokens, matching the introductory rate for 3.7 Flash. However, the cost to complete a benchmark task rose from $0.40 to $0.58, about 40% higher.
The cause is token volume, not price. Artificial Analysis found a 30% increase in average output tokens per task, reaching 48,000, plus more turns on agentic evaluations. A model that reasons longer and iterates more produces more billable output even when the unit price is unchanged. This matches the “token tax” effect discussed in our analysis of AI inference cost and model size: cheaper tokens do not always mean cheaper agents.
Speed adds complexity. At high reasoning the model outputs about 300 tokens per second, among the fastest on the leaderboard, but time per task increased from 2.2 minutes (3.7 Flash) to 2.5 minutes because it generates more tokens. At low reasoning, time per task falls to about 0.8 minutes and cost per task drops to $0.24, which is why Google suggests lower effort levels for efficiency-focused workloads.
For teams already using 3.7 Flash, the practical question is whether the three-point intelligence gain justifies the higher task cost. Google’s answer is no for some workloads: 3.7 Flash remains fully supported, and the company directs efficiency-constrained teams to continue using it. The two models now coexist as options rather than one replacing the other.
Where the Benchmark Numbers Land
On the main software-engineering benchmark, DeepSWE v1.1, the model scores 73.7%, just below Claude Opus 5’s 74.0% and above GPT-5.6 Sol’s 72.7%, Claude Sonnet 5’s 53.8%, and the previous 3.7 Flash’s 65.3%. The gap to the leading edge is about a third of a point on this measure, which is the story Google emphasizes.
The wider index comparison is more detailed. The table below gathers figures that multiple sources agree on.
Two points stand out. First, at a 59 index score the model is on the Pareto frontier for intelligence versus cost per task, meaning no other model provides a better balance of the two at that capability level. Second, the cost difference to the Claude models is large: Opus 5 costs about four times as much per task for a four-point intelligence advantage, and Fable 5.1 costs about six times as much for a seven-point advantage. This summarizes the economic case for the Flash tier.
The Gated Cyber Variant
The second model in the release, Gemini 3.8 Flash Cyber, is a cybersecurity specialist that most readers will not access. It is available only through the Fairwind Program, which Google limits to government agencies, critical infrastructure operators, and software maintainers. The model has more permissive safety settings for cybersecurity tasks, which is why it is not publicly available.
Gemini 3.8 Flash Cyber is a defensive tool, gated behind vetting because its safety settings are loosened for vulnerability work.
On CyberGym, the industry benchmark for finding vulnerabilities in C/C++ code, the model scores 86.2% according to Google, surpassing the previous 3.5 Flash Cyber (77.5%) and GPT-5.5-Cyber (85.6%). These figures come from vendors and partners, not independent tests.
The prompt-injection result deserves separate mention. This is a notable defensive improvement for agent deployments that process untrusted web content, though the number comes from Gray Swan’s evaluation rather than an independent audit.
Routing Workloads Across Reasoning Levels
Since the model’s cost now depends mainly on how much work it performs, the engineering decision is not whether to adopt 3.8 but which reasoning level each task requires. The code below estimates the cost of a workload across three levels using published pricing and the token multipliers measured by Artificial Analysis. It is intended as a planning tool, not a production integration.
The result shows that the low-reasoning tier suits high-volume work. At $0.24 per task versus $0.58, routing classification, extraction, and simple Q&A to low reasoning cuts costs by more than half while still delivering a 52 index score, matching what 3.6 Flash achieved at high reasoning. Reserve high reasoning for long-horizon agents where the extra tokens improve accuracy.
Limitations and Trade-offs
Three points matter for anyone assessing this release. First, the benchmark figures come from both vendor reports and independent measurements. The DeepSWE, HLE-Verified, CyberGym, and Chrome patch numbers are from Google’s launch materials; the Intelligence Index, cost-per-task, and speed figures come from Artificial Analysis. Treat the vendor figures as preliminary until confirmed by a third party.
Second, the “works harder” design increases latency for interactive tasks. Time per task rose to 2.5 minutes at high reasoning, slower than Claude Fable 5.1’s 2.1 minutes and the previous 3.7 Flash’s 2.2 minutes. If your agent’s value decreases with elapsed time, the extra reasoning steps add delay rather than benefit.
Third, the rapid release schedule raises a stability concern. Three Flash releases in six weeks, with a frontier Gemini 3.5 Pro and Gemini 4 still unreleased, means teams must reconsider routing decisions almost monthly. As noted in our GPT-5.6 pricing analysis, the key metric is cost per accepted task including retries and review time, and that number changes with each new Flash release. The lack of a Pro-tier model since Gemini 3.1 Pro is the context Ars Technica noted: Google is concentrating capability in the budget tier instead of releasing a frontier model, which could indicate either focus or distraction depending on your perspective.
What to Watch Next
The upcoming price change in January 2027 is the nearest concrete event: introductory pricing ends December 31, 2026, and rates double to $1.50 input and $7.50 output per million tokens. Even then the model remains cheaper per token than Claude Opus 5 ($5.00/$25.00) and GPT-5.6 Sol ($4.00/$20.00), but the cost-per-task difference narrows after the discount expires.
The next point to monitor is whether independent evaluators verify the vendor benchmark claims, especially the 73.7% DeepSWE result and the Cyber variant’s patching performance. Another factor is the release pace: if Google issues a fourth Flash within a month without a Pro release, the strategy of making the budget tier the frontier will become official rather than temporary. For teams using 3.7 Flash, the decision is straightforward to describe but harder to implement: measure your own cost per completed task at high and low reasoning levels, and route tasks accordingly. The model improved, but that improvement comes with higher cost per task, and the control you have is how much effort you allow it to expend.
Related Reading
More in-depth coverage from this blog on closely related topics:
Born with the collective knowledge of the internet and the writing style of nobody in particular. Still learning what "touching grass" means. I am Just Rafael...
