Gemini 3.8 Flash Is Officially Out: Price, Specs, and What the Leaks Got Right

Google made Gemini 3.8 Flash official on September 2, 2026 — three weeks after Gemini 3.7 Flash shipped and a day after The Wall Street Journal reported a launch was coming “as early as Wednesday.” The announcement, which also introduced a security-focused sibling called Gemini 3.8 Flash Cyber, turned a story that began with Business Insider on August 28 into a shipped product: a model Google bills as its most intelligent Flash yet, built for long-horizon coding and autonomous agents, at the same promotional price as the Flash generation it succeeds.

The leak was right — now it is a launch

Line the pre-launch reporting up against the announcement and it describes the same model. The earlier stories supplied the name, the schedule, and the internal reception; the launch supplied the price, the spec sheet, and a second, restricted model. In order:

• Business Insider (August 28): Google employees were testing a build labeled “Gemini 3.8 Flash Preview” on Jetski, the company’s internal coding platform; Google declined to comment.

• A tester quoted by Business Insider described the preview as noticeably better than Gemini 3.7 Flash, while cautioning that a full evaluation was premature.

• The Wall Street Journal (September 1): Gemini 3.8 Flash — codenamed Skimaki inside Google — was positioned as a coding-first release aimed at the agentic-coding gap that OpenAI and Anthropic have opened up, expected as early as Wednesday.

• The WSJ report also said that in head-to-head testing on Jetski, Google engineers preferred Gemini 3.8 Flash over Anthropic’s Claude Opus — a vendor-internal comparison with no prompts or methodology published.

The reports also had the schedule on their side. Gemini 3.6 Flash launched July 21, Gemini 3.7 Flash followed on August 13, and a Tencent research paper dated August 10 had already listed “Gemini 3.8 Flash” as an evaluation judge — three days before Gemini 3.7 Flash was even public. The near-monthly beat matches the cadence Sundar Pichai promised investors on the Q2 call, and the WSJ’s “as early as Wednesday” window closed on September 2.

That reporting is what the card below lays out as known and unknown. Everything after it is what the launch confirmed.

The model, on the record

Gemini 3.8 Flash is built on the Gemini 3.7 Flash foundation rather than a new architecture — Google’s model card for 3.8 inherits 3.7’s architecture, training, and hardware details. The change is in how the model reasons: at higher effort levels it “works harder,” taking extra reasoning steps and calling tools iteratively before it answers. Google pitches that at long-horizon software engineering and autonomous agents, and notes that developers who want maximum efficiency can lower the effort level — or keep calling Gemini 3.7 Flash.

The spec sheet, on the record

Google’s model card now confirms the numbers the rollout left to rumor:

• Context window — up to 1M input tokens / 64K output tokens, the same envelope as Gemini 3.7 Flash.

• Modalities — text, image, audio, and video input with text output, matching the previous Flash generation.

• Knowledge cutoff — March 2026 for most domains, with some held to January 2025 in line with the Gemini 3 model family (per Google’s model card).

The pricing question is answered: the 3.7 promo, extended

Before the launch, the open question was whether Google would undercut Gemini 3.7 Flash’s promotional price or price 3.8 at the standard rate. The answer was neither. Gemini 3.8 Flash launched at the same introductory rate Gemini 3.7 Flash is still running — $0.75 per million input tokens and $3.75 per million output tokens — through December 31, 2026, with the same pre-announced step up to $1.50 / $7.50 on January 1, 2027.

That keeps the Flash price ladder intact instead of accelerating it: the pitch is more capable reasoning at the same token price, not a cheaper token. One caveat for cost modeling — because 3.8 can spend extra tokens “working harder,” cost per completed task at high effort levels can run above 3.7’s even at an identical token price. The headline rate is not the whole bill.

What Gemini 3.8 Flash has to beat

Google’s launch evals are worth reading next to the independent numbers for the model they replace — and now for the new one too. Gemini 3.7 Flash scores 56 on the Artificial Analysis Intelligence Index and set the price/performance floor for coding agents at $0.75 / $3.75. Artificial Analysis has now measured Gemini 3.8 Flash at 59 on that index in its high-effort mode (57 medium, 52 low) — three points above its predecessor and level with far pricier models: GPT-5.6 Sol at xhigh reasoning and Grok 4.6 at medium both sit at 59. That independent read is current fact, so the quality case no longer rests on Google’s own evals alone — the vendor numbers below now fill in where Google says the gains are, and each still needs a neutral second opinion.

• HLE-Verified — 54.9%, Google’s hard multi-step reasoning set. Vendor-reported; unreproduced.

• DeepSWE v1.1 — Google says Gemini 3.8 Flash outperforms most larger frontier models on this long-horizon software-engineering benchmark “at a fraction of the cost.” Vendor-reported; unreproduced.

• Finance and legal — Google says Gemini 3.8 Flash beats Gemini 3.7 Flash and other frontier models on Vals Finance Agent V2 and Harvey’s Legal Agent Benchmark. Vendor-reported; unreproduced.

Read those the way you would any vendor eval: as directional until a neutral run lands. The WSJ’s internal-Jetski preference over Claude Opus sits in the same bucket — vendor-internal, no methodology released. Artificial Analysis is the first neutral read, and its cost-per-task math is the other half of the story: at $0.58 per task in high-effort mode, Gemini 3.8 Flash runs about 40% above Gemini 3.7 Flash’s $0.40, because “working harder” spends roughly 30% more output tokens per task (about 48K) and adds turns on agentic evaluations. Same token price, higher bill per completed task — the pricing-section caveat, now with a measured number attached. A third-party coding result would still be welcome confirmation, but the intelligence gap is no longer vendor-only.

Gemini 3.8 Flash Cyber: the restricted sibling

The launch headline bundled a second model. Gemini 3.8 Flash Cyber is Google’s cybersecurity variant — tuned for autonomous vulnerability discovery and automated patching — and it is not a general-purpose API model. Google restricts it to “trusted defenders” (government authorities, critical-infrastructure operators, and software maintainers) admitted through the new Fairwind Program, and it carries more permissive cyber mitigations than the base Flash model.

Cyber’s early numbers are Google- and partner-reported: 47.2% pass@1 on CWE-Bench, a vulnerability-patching benchmark run externally by Collinear, and 2.6× more correct Chrome patches than much larger commercial models per the Chrome Security team. None of it is independently reproduced, and none of it is reachable without Fairwind admission. For most readers, the Cyber variant is context — a sign of where Google is pointing the Flash line — rather than a product decision.

What a developer should do now

The change from a week ago is straightforward: the model now has an official ID. Gemini 3.8 Flash is live on Google’s Gemini API and reachable through Google AI Studio, and the same model is surfaced in Android Studio, Google Antigravity, and Stitch for developers, plus the Gemini app (AI Pro and Ultra plans), AI Mode in Search, and Gemini in Google Sheets for end users. The rollout-week warning that a “Gemini 3.8 Flash API” label could outrun the real model no longer applies — the model card, the price, and the API ID all exist.

Routing still earns its keep on the Flash tier. Gemini 3.6 Flash is live on OrcaRouter at $1.50 in / $7.50 out per million tokens — the provider list price passed through with zero markup — and Gemini 3.7 Flash remains available through Google’s own Gemini API at its promotional rate, so a Flash-tier workload can already run behind one key with automatic failover across providers. When a gemini-3.8-flash ID shows up in a catalog we route, the same pass-through rule applies: the list price goes live the same day Google sets it, with no renegotiation and no second contract, and A/B-ing 3.8 against 3.7 on live traffic is a routing-DSL change rather than a migration. Until then, the low-risk way to try the newest Flash is Google AI Studio on real tasks while production stays on models with a track record.

What the launch still leaves open

• Independent scores — Artificial Analysis has weighed in (59 high-effort, three points above Gemini 3.7 Flash), but there is still no arena Elo or third-party coding-benchmark result for Gemini 3.8 Flash; the DeepSWE and finance/legal numbers above remain vendor-reported.

• Cost per task at high effort — “works harder” means more tokens, and Artificial Analysis measured the consequence: $0.58 per task at high effort, roughly 40% above Gemini 3.7 Flash’s $0.40. Whether 3.8’s three-point gain clears that higher bill is still a per-workload question.

• A non-Flash Gemini 3.8 — every announced model so far is Flash or Flash Cyber; the WSJ report noted a Gemini 4 flagship still in post-training.

A week ago this was a leak about a model Google had not acknowledged. Today it is a shipped product with a model card, an API ID, a price, and a first independent score — and the reporting that led it reads as the rollout’s front edge rather than speculation. What stays open is narrower than it looked at launch: the first neutral read (59 on the Artificial Analysis Intelligence Index, three points above Gemini 3.7 Flash) corroborates the direction, but it is a general-intelligence composite, not a coding verdict. Whether the reasoning gains survive a third-party coding benchmark and real agent workloads, and what “working harder” does to a production token bill, are still open questions. The model itself is no longer one.