Gemini 3.8 Flash Explained: Benchmarks, Pricing, Token Use and Score-Version Traps

What Google released

Google released Gemini 3.8 Flash on September 2, 2026 as the stable gemini-3.8-flash model for long-horizon software engineering, autonomous agents and complex enterprise workflows. Google also announced Gemini 3.8 Flash Cyber, a separate restricted configuration for vetted defenders. The two names should not be collapsed: this article evaluates the broadly available Gemini 3.8 Flash unless a Cyber result is explicitly labeled.

Google's current Gemini API documentation lists a 1,048,576-token input limit and 65,536-token output limit. It accepts text, image, video, audio and PDF inputs and returns text. Thinking is available at low, medium and high levels; minimal is not supported. Code execution, function calling, search grounding, structured outputs and caching are supported. Computer use is available in preview, while the Live API, image generation and audio generation are not supported by this stable model.

Primary sources: Google launch announcement and Gemini API model documentation.

Pricing is low per token, but long agent runs can still cost more

Through December 31, 2026, Google lists standard paid API pricing at $0.75 per million input tokens and $3.75 per million output tokens, including thinking tokens. From January 1, 2027, the standard rates are scheduled to rise to $1.50 input and $7.50 output per million tokens. Google also lists context-caching and priority-inference rates separately.

That rate card is only one part of real cost. Google warns that Gemini 3.8 Flash may deliberately use more tokens on difficult, long-running jobs because it can reason in smaller steps, call tools iteratively and verify its own work. A model with the same token price as its predecessor can therefore have a higher cost per completed task if it uses more reasoning tokens or tool turns.

Artificial Analysis made the same point in its September 2 launch evaluation: it reported higher output-token use and more agentic turns than Gemini 3.7 Flash, which raised observed cost per task despite matching token prices.

Primary pricing source: Gemini Developer API pricing. Independent launch analysis: Artificial Analysis — Gemini 3.8 Flash.

Google's benchmark table is not one universal score

Google's evaluation document reports several different benchmarks with different harnesses and provenance. They should not be averaged into one invented ranking.

Google says its own scores are pass@1 unless otherwise noted and that it averages multiple trials for smaller benchmarks to reduce variance. It also says many competitor numbers come from providers' self-reported results unless the methodology names a public leaderboard or another source.

The biggest warning is the dramatic 89.4% on Terminal-Bench 2.1 versus 19.1% on Terminal-Bench 4.0. That is not evidence that the same model suddenly lost 70 points. They are different benchmark versions with different task sets and evaluation conditions. Treating them as interchangeable would be misleading.

Primary methodology: Google DeepMind Gemini 3.8 Flash evaluation methodology.

DeepSWE is not SWE-bench Verified, and neither is SWE-bench Pro

The current Google release package reports DeepSWE v1.1, not SWE-bench Verified. DeepSWE uses its own task set, leaderboard and agent harness. A 73.7% DeepSWE score must not be relabeled as a SWE-bench Verified score.

I also did not find a standardized, source-backed Gemini 3.8 Flash SWE-bench Verified result in the checked Google release materials, nor a reliable SWE-bench Pro result for this exact model. Those fields therefore remain unknown, not zero.

SWE-bench Verified and SWE-bench Pro should also remain separate from one another. They have different datasets, repository mixes and evaluation conditions. If a future result appears, it should be published with the exact benchmark version, harness, task count, reasoning setting, date and source rather than compared loosely with DeepSWE or Terminal-Bench.

HLE-Verified has a disclosed 1,811-item sample

Google reports 54.9% on HLE-Verified. Its methodology says this is accuracy over the full 1,811-item verified set, made up of 668 verified items from the original Humanity's Last Exam set plus 1,143 revised items. Google excludes 689 original items identified as uncertain.

That sample-size disclosure is useful because it makes the headline more interpretable. Google also notes that content-policy filters blocked a significant proportion of questions for one competitor and a small number for another. That is another reason not to interpret a single leaderboard percentage as pure underlying intelligence independent of serving policy.

Artificial Analysis changed the yardstick two days after launch

There is an important benchmark-version trap around Gemini 3.8 Flash.

Artificial Analysis' September 2 launch article reported Gemini 3.8 Flash at 59 on its Intelligence Index and about $0.58 per Index task. On September 4, Artificial Analysis released Intelligence Index v4.2, adding AA-Briefcase and GDP.pdf, removing GPQA Diamond, upgrading graders and rebalancing weights with more private held-out data.

The current Gemini 3.8 Flash high-reasoning model page now shows an Intelligence Index score of 47 and roughly $0.74 per Index task. The current v4.2 methodology contains ten evaluations and estimates the composite's 95% confidence interval at under ±1 based on repeated-evaluation experiments.

This should not be described as a 12-point model regression. The model did not necessarily change; the composite benchmark did. When quoting Artificial Analysis scores, always pin the Index version and observation date.

Sources: Artificial Analysis launch analysis, current Gemini 3.8 Flash model page, and Intelligence Index v4.2 methodology.

Latency is still under-specified for a practical buying decision

Google markets Gemini 3.8 Flash as retaining Flash-level speed and says it has the same speed and low cost positioning as 3.7 Flash, but the checked first-party pages do not publish one representative numeric latency figure that applies across reasoning levels, prompt lengths, tool calls and regions.

Artificial Analysis' current model page also does not expose a stable output-speed number for the high-reasoning entry in the text-accessible result used for this verification pass. That means a responsible comparison should leave representative latency and tokens-per-second unknown rather than import an old measurement from a different reasoning setting or benchmark version.

For production work, measure time-to-first-token, output tokens per second, end-to-end task time, total tokens, tool turns and retry rate on the exact API tier and reasoning level you plan to deploy.

Model-card limitations matter for agents

Google's model card lists hallucinations, occasional slowness or timeouts and increased token use at higher effort as known limitations. It gives a March 2026 headline knowledge cutoff while noting that effective freshness varies by domain and that some areas may resemble the older Gemini 3 family knowledge boundary.

Those caveats are especially relevant for long-horizon agents. More tool use and self-verification can improve completion quality, but they also increase runtime, token consumption and the number of places where an external tool or stale assumption can fail. The correct optimization target is therefore not only benchmark pass rate; it is often cost and latency per successfully completed workflow.

Primary source: Gemini 3.8 Flash model card.

Early public feedback is mixed and highly selection-biased

Early Reddit discussions contain both positive and negative anecdotes. One pre-release/early-access poster reported strong prompt adherence and a successful one-try visual-generation-style test compared with an earlier Flash version. Other users praised the model's willingness to think through tasks rather than stop early.

Negative reports include an example in which a user said Gemini supplied fabricated social-media links while attempting a research-style query, as well as comments complaining about instruction-following and unexpectedly fast subscription-limit consumption after the rollout.

These posts are useful for identifying failure modes to test—citation verification, instruction adherence, token use and product limits—but they are not controlled experiments and they come from self-selected users on unknown product routes, system prompts and reasoning settings. They should not be converted into a claim that the community broadly loves or dislikes the model.

Discussion examples: early positive test, hallucinated-link report, and usage-limit discussion.

No sufficiently attributable first-hand X post with reproducible performance evidence was verified in this run, so this article does not invent an X consensus.

Practical takeaway

Gemini 3.8 Flash is a broadly available, 1M-context multimodal model with competitive introductory token pricing and a strong focus on agentic software and enterprise workflows. Its most useful evidence is not one headline number but the combination of: Google's disclosed per-benchmark harness notes, the 1,811-item HLE-Verified evaluation, third-party cost-per-task measurements, and clear documentation that higher reasoning can use more tokens.

For a fair bake-off, pin the model ID and reasoning level, then run the same tasks with the same tools, timeout, retry policy and budget. Keep DeepSWE, Terminal-Bench 2.1, Terminal-Bench 4.0, SWE-bench Verified and SWE-bench Pro in separate columns. Track both success rate and cost/latency per successful task. Finally, record the benchmark version: the Artificial Analysis shift from its September 2 launch score to v4.2 two days later shows how quickly a leaderboard number can change even when the model itself has not.

Confidence is high on model identity, context/output limits, pricing schedule and Google's disclosed benchmark methodology because those come from current primary documentation. Confidence is medium on cross-vendor ranking because several comparison rows use different source provenance or public leaderboards. Confidence is low on broad user sentiment because current public feedback is sparse, self-selected and configuration-dependent.