Gemini 3.8 Flash Ships at $0.75 With Fairwind Cyber Gate

Google shipped Gemini 3.8 Flash on September 2, 2026 as the third Flash-line release in roughly six weeks, with an intro list price of $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026. The model id is gemini-3.8-flash. The same day, Google opened Gemini 3.8 Flash Cyber behind the Fairwind Program for trusted defenders. Public API buyers get a cheap multimodal workhorse. Security teams get a gated cyber stack with no public Cyber API price sheet.

That split is the story. Flash is for shipping agents and long-context jobs. Cyber is for people Google already trusts to touch vulnerability work. If you mix those audiences in one procurement ticket, you will either overpay or hit a wall.

What Gemini 3.8 Flash Is

Gemini 3.8 Flash is Google's low-latency multimodal Flash model for text, image, audio, and video, with a 1 million token input context and up to 64K output tokens. It is built for high-volume agent loops, document-heavy RAG, and multimodal triage where Sol-class or Opus-class spend is wasteful.

The cadence matters as much as the card. Three Flash drops in about six weeks means Google is iterating the cheap tier the way cloud providers iterate SKUs: keep the model id family familiar, change the number, reset the FinOps spreadsheet. Treat gemini-3.8-flash as a pin-able id for Q4 2026, not a forever alias.

Pricing Through Dec 31 and the Jan 1 Jump

Intro pricing is $0.75 / $3.75 (input / output per million tokens) until December 31, 2026. From January 1, 2027, Google lists $1.50 / $7.50. That is a clean 2x step on both sides.

For teams running multi-million-token agent sessions, the intro window is a real discount, not a rounding error. A coding harness that burns 200M input and 40M output tokens in a month pays about $150 + $150 = $300 at intro rates, versus $300 + $300 = $600 after the cliff. Budget 2027 at the higher band even if you ship against intro pricing today. Track live rates on the LLM API Pricing Tracker rather than freezing a screenshot from launch day.

Context, Modalities, and Where Flash Fits

The 1M input window is the product. You can park large repos, ticket dumps, design PDFs, and short video clips in one call instead of building a brittle chunking pipeline. The 64K output cap is enough for long patches and multi-file proposals, but it is not an unlimited dump. If your agent needs novel-length generation, plan multi-turn continuation.

Modality coverage (text, image, audio, video) makes Flash a strong default for support desks, QA bots that watch screen recordings, and infra bots that ingest dashboards as images. It does not automatically mean you should replace a coding-primary model for every Cursor or Claude Code session. For peer comparisons across the September stack, see the Astra vs Fable 5.1 vs Gemini 3.8 Flash guide and the broader Best AI Models 2026 hub.

What Fairwind and Gemini 3.8 Flash Cyber Are

Gemini 3.8 Flash Cyber is the cyber-capable sibling distributed through Google's Fairwind Program for trusted defenders: government teams, critical infrastructure operators, and software maintainers. At launch Google cited 650+ Fairwind participants, with named references including Snowflake, CrowdStrike, and Datadog.

There is no public Cyber API and no public Cyber price list. If you are a solo developer or a product team without a Fairwind seat, you do not get Cyber capabilities by flipping a model id. That is intentional. The public Flash card stays the shipping surface; Cyber stays in a vetted channel.

CyberGym Scores and CodeMender Claims

On Google's reported CyberGym numbers, Gemini 3.8 Flash Cyber lands near ~86.2%, ahead of Mythos 5 at ~83.8% and GPT-5.6 Sol at ~83.6%. Treat those as vendor-reported benchmark points, not an independent red-team grade for your environment.

Google also describes a CodeMender harness aimed at find / verify / patch loops, and claims Chrome-related patches are about 2.6x more accurate under that harness versus prior baselines they measured. Useful as a product signal. Not a substitute for your own staging tests, fuzz corpus, and human review before any patch lands in production.

Our Analysis: How Developers Should Buy Flash vs Fairwind

Public Flash is a FinOps win for multimodal and long-context work through year-end. Fairwind Cyber is an access-control product more than a price product. Most teams need both strategies on paper and only one in the API key vault.

Practical split:

1. Default agents to public gemini-3.8-flash. Pin the id. Cap monthly spend. Rehearse the Jan 1, 2027 2x price jump in your forecast.
2. Do not invent a Cyber endpoint. If your security org is not in Fairwind, plan remediation with public models plus traditional scanners. Waiting for a gray-market Cyber key is a compliance failure waiting to happen.
3. Apply if you are a maintainer or critical-infra team. Fairwind's cited participant mix (cloud data platforms, endpoint vendors, observability) shows Google wants production defenders, not curiosity accounts.
4. Keep harnesses portable. Agent skill packaging from Agent Plugins 1.0 matters more than loyalty to one Flash number. When 3.9 lands, you swap the id, not the whole pipeline.
5. Separate coding SOTA from cheap multimodal. Use Flash for volume and context. Keep Astra / Fable-class models for computer-use and coding benches where they still win on the public cards.

China and Bing search traffic already hunt Gemini Flash pricing and self-host alternatives. Number-led titles and a clear Fairwind gate explanation serve that audience without promising Cyber access you cannot sell.

FinOps Worked Examples for Flash

Here is a concrete monthly sketch for a mid-size product team. Assume 120M input tokens and 25M output tokens of Flash traffic across agents, support summarization, and nightly multimodal jobs.

That is still cheap versus a Standard $10/$50 coding model on the same volume ($1,200 + $1,250 = $2,450). The trap is mixing workloads. If you put computer-use agents and IDE coding onto Flash because it is inexpensive, you save money and burn quality. Keep Flash on the jobs that match its strengths: long context, multimodal intake, and high-QPS summarization.

Also model cache behavior. Flash intro pricing does not automatically copy Anthropic's cache-read discounts. If your prompts are huge and repetitive, Fable 5.1 may still win on effective dollars even when Flash wins on sticker price. Put both through the LLM API Pricing Tracker with your real cache hit rate.

Migration Notes From Earlier Flash Builds

If you are on Gemini 3.x Flash from the prior five-to-six week window, treat 3.8 as a pin change plus a regression pass, not a rewrite. Re-run your top 20 agent traces. Check tool-call JSON validity, multimodal token inflation on video, and output truncation near 64K. Update eval baselines so a silent quality drop does not hide behind "we upgraded Flash."

Teams that abstract model ids behind an env var will finish in an afternoon. Teams that hardcode ids in six services will invent a week of chaos. Portable skill packaging from Agent Plugins 1.0 is still the cheapest insurance against Google's Flash cadence.

What To Watch Next

Watch three clocks: the Dec 31 intro expiry, Fairwind expansion beyond the launch 650+, and whether Google ever lists a public Cyber SKU (unlikely soon). Also watch whether CodeMender-style patch loops show up in third-party harnesses with reproducible diffs. Until then, buy Flash as a commodity multimodal engine and treat Cyber as a membership, not a model dropdown.

Key Takeaways

• Gemini 3.8 Flash launched Sept 2, 2026 as the third Flash in ~6 weeks; model id gemini-3.8-flash
• Intro price $0.75 / $3.75 per MTok input/output through Dec 31, 2026; $1.50 / $7.50 from Jan 1, 2027
• 1M input context, 64K max output; text, image, audio, video
• Gemini 3.8 Flash Cyber ships only via Fairwind (650+ orgs at launch; Snowflake, CrowdStrike, Datadog cited)
• CyberGym ~86.2% vs Mythos 5 ~83.8% and GPT-5.6 Sol ~83.6% (vendor-reported)
• CodeMender find/verify/patch harness; Google claims ~2.6x more accurate Chrome patches in their tests
• No public Cyber API or Cyber pricing
• For developers: pin public Flash for volume; apply Fairwind only if you are a trusted defender; budget the 2027 2x cliff now
• What to watch: intro expiry, Fairwind growth, and any future public Cyber SKU

Sources

• Google Gemini 3.8 Flash and Fairwind Program announcements (Sept 2, 2026)
• Google-stated intro and post-intro list prices for gemini-3.8-flash
• Google-reported CyberGym comparisons and CodeMender / Chrome patch accuracy claims (Sept 2026)
• Fairwind launch participant count and cited customers (Snowflake, CrowdStrike, Datadog)