Gemini 3.8 Flash is generally available as gemini-3.8-flash since 2 September 2026 at $0.75 per million input tokens and $3.75 per million output — exactly 3.7 Flash's introductory price — and both numbers double to $1.50 / $7.50 on 1 January 2027. It beats 3.7 Flash on every benchmark Google published, most sharply on Terminal-Bench 2.1 (90.8% vs 81.6%), and beats GPT-5.6 Terra (87.4%) and Claude Sonnet 5 (80.4%) on the same test. The catch is in the model card: 3.8 is built on 3.7 rather than a new base, it "works harder" by spending more thinking tokens, and Google tells efficiency-first workloads to stay on 3.7.
That last sentence is unusual for a launch, and it is the reason this guide is a decision table rather than a migration checklist. Sources: Google's model page and the DataCamp and eesel breakdowns published on launch day.
Pricing, and the 1 January cliff
| Period | Input / MTok | Output / MTok (includes thinking) |
|---|
| Now to 31 December 2026 | **$0.75** | **$3.75** |
| From 1 January 2027 | $1.50 | $7.50 |
Batch and Flex are half those rates; Priority is 1.8×. Multimodal input — text, image, audio, video, PDF — sits inside the same 1M-token window with 64k output. Three things follow from the table. Output pricing includes thinking tokens, and 3.8 thinks more than 3.7 at the same effort setting, so the per-request cost is higher even at identical list prices. Any budget you approve on September pricing is wrong by 2× in four months. And Gemini remains an order of magnitude below the frontier tier: Claude Fable 5.1 and GPT-6 Astra are $10 / $50. Thirteen Gemini 3.8 Flash requests cost roughly one Astra request at the same token counts — before Astra's cache discount and before Gemini's extra thinking, which is why you should model both in the AI model cost calculator rather than eyeballing it.
Benchmarks: a coding release, not a reasoning release
| Benchmark | Gemini 3.8 Flash | Gemini 3.7 Flash | Delta |
|---|
| Terminal-Bench 2.1 | **90.8%** | 81.6% | +9.2 |
| SWE-Bench Pro | 61.6% | 60.4% | +1.2 |
| SWE-Atlas | 51.9% | 48.0% | +3.9 |
| τ³-bench Banking (tool use) | 38.1% | 30.9% | +7.2 |
| CharXiv (multimodal reasoning) | 86.2% | 84.5% | +1.7 |
| Humanity's Last Exam | 45.4% | 45.7% | −0.3 |
| HLE-Verified | 54.9% | — | — |
The shape is unmistakable. Terminal work and multi-step tool use jump; general reasoning is flat (HLE actually dips a tenth of a point). Google also says 3.8 Flash outperforms "most larger frontier models" on DeepSWE v1.1 for long-horizon coding without publishing the percentage, and that it beats Claude Opus 5 on three benchmarks. Against the competition on Terminal-Bench 2.1, 3.8 Flash's 90.8% sits above GPT-5.6 Terra at 87.4% and Claude Sonnet 5 at 80.4%.
If you have read our Gemini 3.5 Flash guide from May, note the velocity: 3.5 Flash scored 76.2% on Terminal-Bench 2.1 four months ago. The Flash line has gained 14.6 points on that benchmark in one summer, across three releases (3.6 on 21 July, 3.7 on 13 August, 3.8 on 2 September).
Where the extra score comes from — and what it costs
Google is explicit that 3.8 is 3.7 with more deliberate reasoning at higher effort levels, not a new pretraining run. That explains both the profile (tool-use and terminal tasks reward extra planning steps; trivia does not) and the cost warning. In practice: at the same effort setting, expect more thinking tokens per response than 3.7, and because thinking is billed as output at $3.75 per million, the per-task cost rises even though the price list is unchanged.
The right experiment is not "is 3.8 better" — it is — but "is 3.8 at a lower effort setting better than 3.7 at a higher one for the same spend". Run twenty of your real tasks through both, log output token counts, and compare pass rate per dollar. The token counter helps you size the inputs; the output side you have to measure.
Decision table
| Your workload | Recommendation | Why |
|---|
| Coding agents, terminal automation, long-horizon repo tasks | Migrate to 3.8 | +9.2 on Terminal-Bench, +7.2 on τ³-bench; the extra thinking is exactly what these tasks need |
| Customer-support chat, classification, extraction at volume | Stay on 3.7 | Google's own guidance; reasoning gains are minimal and thinking tokens raise cost |
| Document / chart understanding | Test first | +1.7 on CharXiv is real but small; measure output tokens |
| Anything with a January 2027 budget | Re-forecast now | 2× price on 1 Jan regardless of which Flash you run |
| Vulnerability discovery, security tooling | Apply to Fairwind | The Cyber variant is not on the public API |
The Cyber variant
Gemini 3.8 Flash Cyber shares the base model but is tuned and restricted for defenders. It is available only through Google's Fairwind programme — 650-plus partners including CrowdStrike, Datadog, Palo Alto Networks, Menlo Security and Snowflake, plus governments, healthcare, telecom and Google Cloud customers who apply. Google claims a 70%-plus real-world vulnerability discovery rate across 20 programming languages and 47.2% pass@1 on CWE-Bench, a hair behind the leading frontier model at 47.8%, with the emphasis on patching over exploitation. Google's framing is that it is the most capable cybersecurity model available, surpassing Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol on autonomous vulnerability discovery.
Three labs shipped gated cyber tiers in 72 hours this week; the pattern and what it means for teams building security tooling is in our cyber-capable model guide.
Migration mechanics
Change the model string to gemini-3.8-flash; the API surface is unchanged from 3.7. Keep your effort setting where it is for the first run and record output tokens per request, because that is the number that moves. If you use Batch or Flex for anything asynchronous, the half-price tiers make 3.8 cheaper than 3.7 on-demand — a nightly repo-maintenance agent belongs there. And set a calendar reminder for December: every Flash budget written this quarter needs a 2× line for January.
Quick answers
What is the Gemini 3.8 Flash model ID?
gemini-3.8-flash, available through Google AI Studio and the Gemini Enterprise Agent Platform since 2 September 2026. The request format is unchanged from 3.7 Flash.
How much does Gemini 3.8 Flash cost?
$0.75 per million input tokens and $3.75 per million output through 31 December 2026, rising to $1.50 and $7.50 on 1 January 2027. Output pricing includes thinking tokens; Batch and Flex are half price and Priority is 1.8×.
Should I upgrade from Gemini 3.7 Flash?
For coding agents, terminal automation and multi-step tool use, yes — the gains are 7 to 9 points on the relevant benchmarks. For high-volume chat, classification and extraction, Google itself says to stay on 3.7, because 3.8 spends more thinking tokens for little reasoning gain.
What is Gemini 3.8 Flash Cyber?
A variant tuned and restricted for defensive security work, available only through Google's Fairwind programme of 650-plus partner organisations. It shares the base model and reports a 70%-plus real-world vulnerability discovery rate across 20 languages.
How does it compare with Claude Sonnet 5?
On Terminal-Bench 2.1, 3.8 Flash scores 90.8% against Sonnet 5's 80.4%, at $0.75 input versus Sonnet's $2. Sonnet 5 keeps a stronger general-reasoning reputation; for terminal and agent work the Gemini number is hard to argue with until January's price doubling narrows the gap.
For routing across Gemini, Claude and OpenAI in one codebase, the Claude Code Production Pack ships the config-driven model-routing rules we use, and the Agent Prompt Vault has the evaluation prompts for the twenty-task comparison above. Every product mentioned is available at wowhow.cloud — pay once, ship forever.
Originally published at wowhow.cloud
Top comments (0)