Gemini 3.8 Flash is no longer a rumor. Google officially confirmed the model on September 2, 2026, alongside a new cybersecurity-focused sibling called Gemini 3.8 Flash Cyber, in a joint launch post from Google’s Tulsee Doshi and Raluca Ada Popa. The confirmation landed a day earlier than the “Wednesday” launch date that had been circulating in leaks, closing the book on weeks of speculation — and notably, it makes Gemini 3.8 Flash the fourth distinct Flash model Google has shipped in under four months, an iteration pace that has no real precedent in the Flash line’s history. But the more interesting story is still the one underneath it: Google, OpenAI, and Anthropic are locked in a token-pricing fight that has already cut small-model costs by as much as 80% in five weeks, and Gemini 3.8 Flash’s steady pricing is the latest data point in that game.
Before that confirmation, the model spent weeks in leak limbo. According to reporting from Shattered.io, no public changelog, benchmark page, or pricing sheet for a model called Gemini 3.8 Flash existed as of August 28, 2026. Internally, the project reportedly carried the code name “Skimaki,” and the earliest public trace of it was a Tencent research paper dated August 10, 2026 — weeks before Google said anything. The name also surfaced in a Business Insider report describing an internal preview tested by Google employees on Jetski, an internal coding platform. That gap between what leaked for weeks and what Google eventually confirmed is the real news here, and it says a lot about how fast the AI pricing war has gotten.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What’s Actually Confirmed About Gemini 3.8 Flash
Strip away the speculation and the confirmed facts are now substantial. Google officially confirmed Gemini 3.8 Flash on September 2, 2026, in a joint launch post from Tulsee Doshi and Raluca Ada Popa, alongside a dedicated security-focused variant, Gemini 3.8 Flash Cyber. That announcement closed the gap that research firm LuminaBench first flagged on August 28, 2026, when it reported the internal preview via a summary published by kie.ai — at that point, Google had published no release date, pricing, context-window specification, benchmark score, public API endpoint, or license for the model. Less than a week later, all of that had an answer.
Pricing and specs are confirmed too. Google’s own blog lists Gemini 3.8 Flash at $0.75 per million input tokens and $3.75 per million output tokens as of September 2026 — the same rate Gemini 3.7 Flash launched with — though third-party tracker ArtificialAnalysis.ai lists a standard rate of $1.50 per million input tokens and $7.50 per million output tokens for that same month, with cached input tokens discounted a further 90% off the standard input rate. Either way, the model ships with a confirmed 1,048,576-token context window (roughly 1 million tokens) and a maximum of 65,536 output tokens, per MarkTechPost’s September 2026 reporting. That introductory rate isn’t permanent, either: Google has disclosed that pricing rises to $1.50 per million input tokens and $7.50 per million output tokens on January 1, 2027, effectively doubling the cost of the model once the promotional window closes. NokiaPowerUser had reported that community and developer leaks placed the new build in internal testing less than three weeks after Google rolled out Gemini 3.7 Flash, which itself only shipped in mid-August 2026 — and that compressed cadence turned out to be accurate. From internal testing to a confirmed public launch took just over three weeks.
Gemini 3.8 Flash Cyber: Google’s New Security-Focused Variant
Alongside the base model, Google used the September 2 announcement to introduce Gemini 3.8 Flash Cyber, a variant tuned for cybersecurity workloads. The joint post from Tulsee Doshi, who leads Gemini model development, and Raluca Ada Popa, who leads Google’s security and privacy research, frames the Cyber variant as purpose-built for tasks like vulnerability triage, security-log analysis, and defensive tooling rather than general-purpose chat or coding. Google’s launch post did not break out separate pricing for the Cyber variant, so it is not yet clear whether it will carry the same $0.75/$3.75 per million token rate as standard Gemini 3.8 Flash or a distinct enterprise-security price tier down the line.
The pairing signals that Google is willing to ship a security-specific model in the same breath as a mainstream Flash release rather than holding it back as a separate premium product. That is a notable strategic choice: pairing a security-focused launch with Raluca Ada Popa’s team gives Gemini 3.8 Flash Cyber an immediate credibility angle with enterprise security buyers that a generic model announcement would not carry on its own.
Before the Confirmation: What Early Testers Were Saying
The qualitative picture from leak reporting was more consistent than the technical one turned out to be. NokiaPowerUser’s sourcing described the build as targeting “reduced slop” (fewer hallucinated or off-target outputs), improved agentic logic for tool use and multi-step planning, and faster coding responses compared with Gemini 3.7 Flash. Shattered.io’s reporting added a specific data point: one early tester described the preview as “noticeably better than 3.7 Flash,” though that impression came from informal use rather than a benchmark run. That informal read has since picked up quantitative backing: ArtificialAnalysis.ai’s September 2026 benchmark run rates Gemini 3.8 Flash “high” on its intelligence index, with a score of 41, at least partially validating the pre-launch chatter without settling every capability claim leak reporting made. The general-availability release also confirmed a capability the leaks never mentioned: developers can now choose between three tunable thinking levels — low, medium, and high — letting teams trade latency against reasoning depth on a per-request basis rather than accepting a single fixed behavior.
A 1-million-token context window had circulated in community posts tied to the leak well before Google confirmed it. Pricing was a similar story at the time — undisclosed, with Gemini 3.7 Flash’s own introductory rate of $0.75 per million input tokens and $3.75 per million output tokens, reported by kie.ai, serving as the only real anchor point. That anchor turned out to be close to where Gemini 3.8 Flash landed: Google’s blog carried the $0.75/$3.75 rate forward, and by September 2026 Google’s own API changelog recorded Gemini 3.8 Flash as generally available under the model ID gemini-3.8-flash, with Google’s Interactions API documentation separately confirming GA status alongside the 1-million-token context window and a 64k max output token limit.
Why Google’s Confirmation Changes the Calculus
For weeks, Shattered.io’s reporting turned on one point that was easy to miss in the rush to cover a “release date”: Google had not issued any public statement confirming or denying a model under the Gemini 3.8 Flash name. That silence did a lot of work in this story. Every spec that circulated before September 2 — the context window claim, the pricing, even the “Skimaki” code name — originated from leaks, community posts, and a single Business Insider report about internal Jetski testing, not from a Google blog post or executive comment. The September 2 launch post from Tulsee Doshi and Raluca Ada Popa changed that in one move, converting every one of those leaked data points into confirmed fact at once.
That distinction mattered for anyone making a purchasing or migration decision in the days before the announcement. Treating an unconfirmed internal preview as a locked release date would have meant planning around a number Google had not put on paper. Now that distinction is moot: Gemini 3.8 Flash and Gemini 3.8 Flash Cyber are both officially live, with Google’s own pricing and specifications replacing every leaked estimate.
The Real Story: 2026’s AI Token Price War
While Gemini 3.8 Flash sat in leak limbo, the pricing war around it was extremely well documented. On July 30, 2026, OpenAI announced sweeping cuts to its GPT-5.6 line, according to CNBC and an OpenAI community announcement. GPT-5.6 Luna, the smallest model in the family, dropped 80% to $0.20 per million input tokens and $1.20 per million output tokens. GPT-5.6 Terra, the mid-tier model, fell 20% to $2.00 per million input tokens and $12.00 per million output tokens. GPT-5.6 Sol, OpenAI’s flagship, kept its price near $5 input and $30 output per million tokens.
OpenAI framed the cuts as a response to businesses scrutinizing AI spend more closely as usage scales, which is corporate shorthand for a fight over developer default choices. Every point along OpenAI’s pricing ladder from GPT-5 nano at roughly $0.05 per million input tokens up to Sol at $30 per million output tokens now has a Google or Anthropic equivalent sitting nearby, and that’s precisely the market Gemini 3.8 Flash has now entered without needing a price change to stay competitive.
Anthropic’s Reversal Shows How Price-Sensitive This Market Has Become
Anthropic’s own pricing history this year underlines the same pressure. Claude Sonnet 5 launched at a promotional $2 per million input tokens and $10 per million output tokens, with a standard price of $3/$15 originally scheduled to take effect September 1, 2026 — a 50% increase. Anthropic later cancelled that planned hike and made the launch rate the permanent price, keeping Sonnet 5 at $2/$10. That’s a company walking back a scheduled price increase rather than risk losing developers to cheaper competitors, and it’s the clearest signal yet of how sensitive AI buyers have become to per-token costs.
Further up Anthropic’s stack, Claude Opus 5 sits around $5 input and $25 output per million tokens, while Claude Fable 5, the company’s most capable tier, runs closer to $10 input and $50 output per million tokens. Compared to OpenAI’s Sol at roughly $5/$30, Opus 5 undercuts on output pricing by close to 17% while matching on input — a small but real edge Anthropic has leaned on in its enterprise sales pitch, according to pricing trackers cited by Opslyft’s 2026 pricing comparison.
Where Gemini’s Flash Line Fits in the Price Ladder
Google’s existing Flash tier is already the budget anchor most developers compare everything else against. Gemini 2.5 Flash-Lite runs $0.075 per million input tokens and $0.30 per million output tokens with a 1-million-token context window, undercutting even OpenAI’s post-cut Luna pricing. Gemini 2.5 Flash sits at roughly $0.25 input and $1.50 output per million tokens, also with a 1-million-token window. Gemini 3.7 Flash, the model that Gemini 3.8 Flash has now replaced, introduced at $0.75/$3.75 per million tokens according to kie.ai’s reporting, and Google carried that exact rate forward to 3.8 Flash rather than raising it — our own coverage of that Gemini 3.7 Flash cost cut tracked the same downward pressure across Google’s lineup.
One tier up, Gemini 3.1 Pro is priced at $2 input and $12 output per million tokens — a near-exact match for OpenAI’s post-cut GPT-5.6 Terra. That convergence isn’t a coincidence; it’s what a genuine price war looks like when none of the three labs wants to be visibly more expensive than the other two at any given capability tier.
| Model | Provider | Input ($/M tokens) | Output ($/M tokens) | Context Window | Status |
|---|---|---|---|---|---|
| Gemini 2.5 Flash-Lite | $0.075 | $0.30 | 1M tokens | Generally available | |
| GPT-5.6 Luna | OpenAI | $0.20 | $1.20 | Not disclosed | Generally available (post-cut, July 30, 2026) |
| Gemini 2.5 Flash | $0.25 | $1.50 | 1M tokens | Generally available | |
| Gemini 3.7 Flash | $0.75 | $3.75 | Not officially specified | Generally available | |
| Claude Sonnet 5 | Anthropic | $2.00 | $10.00 | Not disclosed | Generally available (launch price made permanent) |
| Gemini 3.8 Flash | $0.75 | $3.75 | 1M tokens | Generally available (confirmed Sept 2, 2026) | |
| Gemini 3.8 Flash Cyber | Not separately disclosed | Not separately disclosed | 1M tokens | Generally available (confirmed Sept 2, 2026) |
Mid-Tier and Flagship Pricing: The Bigger Battle
Zoom out to the mid-tier and flagship models and the same convergence pattern holds. GPT-5.6 Terra and Gemini 3.1 Pro are now priced identically at $2/$12 per million tokens. At the top end, GPT-5.6 Sol (roughly $5/$30) and Claude Opus 5 (roughly $5/$25) sit within a few dollars of each other, with Anthropic’s slightly cheaper output pricing the main differentiator. Claude Fable 5, Anthropic’s premium tier at around $10/$50 per million tokens, remains the priciest option among the three labs’ publicly listed models, positioned for workloads where raw capability outweighs cost per token.
| Model | Provider | Input ($/M tokens) | Output ($/M tokens) | Tier Notes |
|---|---|---|---|---|
| GPT-5.6 Terra | OpenAI | $2.00 | $12.00 | Cut 20% on July 30, 2026 |
| Gemini 3.1 Pro | $2.00 | $12.00 | Matches Terra exactly | |
| Claude Opus 5 | Anthropic | ~$5.00 | ~$25.00 | ~17% cheaper output than Sol |
| GPT-5.6 Sol | OpenAI | ~$5.00 | ~$30.00 | Flagship, price held steady |
| Claude Fable 5 | Anthropic | ~$10.00 | ~$50.00 | Anthropic’s top capability tier |
For a closer breakdown of that flagship tier, our earlier coverage of Claude Opus 5 vs. Mythos 5 vs. Gemini 3.1 Pro and GPT-5.6 Sol vs. Terra vs. Luna both map onto the same pricing table shown above, just from opposite ends of the fight.
Historical Context: How Google’s Flash Line Got Here
Flash has never been Google’s showcase model — that’s the point of it. Since its debut, the line has existed to undercut rivals on cost per token while staying usable for real production workloads like chat support, summarization, and lightweight coding assistants. Gemini 2.5 Flash-Lite’s $0.075/$0.30 pricing with a 1-million-token window set an aggressive low bar that neither OpenAI nor Anthropic matched outright until this summer’s cuts. Gemini 3.7 Flash, shipping in mid-August 2026, pushed capability up while keeping pricing well below Google’s own Pro tier. Google DeepMind’s own model card, published in September 2026, formally describes Gemini 3.8 Flash as the next iteration over Gemini 3.7 Flash — corporate confirmation of the version-to-version relationship that leak reporting had only assumed for weeks.
What was different about the Gemini 3.8 Flash cycle, based on NokiaPowerUser’s and Shattered.io’s sourcing, was speed: internal testing began within three weeks of 3.7 Flash’s public rollout, and Google confirmed the finished model on September 2, 2026 — a little over three weeks after that testing reportedly started. Historically, Google had spaced major Flash revisions months apart, not weeks. That compressed cadence held, and it suggests Google’s AI teams are treating iteration speed itself as a competitive weapon, not just price or benchmark scores.
Market Impact: What Faster Price Cuts Mean for Enterprise AI Budgets
For engineering teams running production AI workloads, the practical impact of this price war shows up directly on the invoice. A team processing a billion output tokens a month on GPT-5.6 Luna now pays roughly $1,200 instead of $6,000 before the July 30 cut — a difference that reshapes the economics of high-volume use cases like customer support bots, content moderation, and batch document processing. Multiply that across Gemini’s Flash tiers and Claude Sonnet 5’s frozen pricing, and the net effect is that budget-tier AI has gotten meaningfully cheaper twice in the past six weeks without any team having to switch providers. That cost math now extends to agent workloads too: Google’s Gemini Enterprise Agent Platform release notes, dated September 2026, mark Gemini 3.8 Flash as generally available and production-ready, meaning teams building agentic pipelines can adopt the model without waiting on a separate enterprise rollout.
That volatility cuts both ways for procurement teams. Locking into annual enterprise contracts at today’s rates risks paying a premium if another 20–80% cut lands in the next quarter, which is exactly what happened to teams that signed GPT-5.6 agreements before the July cuts. The safer near-term move, several pricing analysts note, is shorter commitment terms or usage-based contracts that let teams capture price drops as they land rather than get stuck above market rate.
Competitive Comparison: Three Labs, Three Different Strategies
The three labs are not fighting this price war the same way. OpenAI is cutting hardest at the bottom of its lineup (Luna’s 80% reduction) while holding its flagship Sol steady, a strategy aimed at maximizing developer adoption of cheap, high-volume endpoints. Anthropic is defending its mid-tier position by freezing Sonnet 5 rather than raising it, prioritizing retention over margin expansion. Google, now that Gemini 3.8 Flash and Gemini 3.8 Flash Cyber are confirmed, is betting on speed of iteration rather than headline price cuts — it shipped a materially improved Flash model plus a new security-focused variant within weeks of the last one, at the same price, rather than waiting for a major version jump to justify a new price point.
Each approach reflects each company’s existing strengths. OpenAI has the largest developer base to defend with aggressive small-model pricing. Anthropic has built its enterprise reputation on reliability and safety framing more than raw price. Google has the infrastructure scale — its own TPUs — to subsidize rapid, low-cost Flash iterations that would be harder for a lab renting third-party GPU capacity to match. Readers curious how a fourth open-weight contender factors into this fight can see our coverage comparing DeepSeek V4-Flash, Gemini 3.7 Flash, and Qwen3.8-Flash-Next, where price gaps run even wider than among the three US labs.
The GOOGL Angle: What’s Verified and What Isn’t
Social and market chatter tied the earlier Gemini 3.8 Flash leak to a positive move in Alphabet’s after-hours share price, and outlets including TradingView, TipRanks, and Crypto Briefing were cited circulating that narrative. At the time, what was harder to pin down independently was a verified, timestamped price move explicitly attributed to the leak rather than to broader market conditions on the day — none of the sourced technical reporting on the leak tied a specific stock percentage to the story, and Google itself had made no statement that would give Wall Street a confirmed catalyst to price in. That changed on September 2, 2026, when Google’s launch post gave the market an actual, confirmed catalyst for the first time, though no verified post-announcement stock figures are available as of this writing.
That distinction is worth holding onto: a leak-driven rumor and a confirmed product launch move markets very differently, and treating speculative chatter as a hard catalyst is how this kind of story gets overstated. Readers who want the market-reaction angle in more depth can see our separate report on the GOOGL after-hours move tied to the original release rumor.
Predictions: Where This Goes From Here
- Gemini 3.8 Flash Cyber gets its own disclosed pricing within the next few weeks, once Google has usage data to justify a security-specific tier separate from the base model’s $0.75/$3.75 rate.
- OpenAI or Anthropic responds with another small-model price adjustment within 60 days of the September 2, 2026 Gemini 3.8 Flash launch, extending the tit-for-tat pattern seen since July 2026.
- Third-party benchmarks of Gemini 3.8 Flash’s “reduced slop” and agentic-coding claims arrive within weeks, testing whether the qualitative gains early testers described hold up outside informal use.
- Enterprise security teams become the first serious adopters of Gemini 3.8 Flash Cyber, given its launch alongside Raluca Ada Popa’s security and privacy research group rather than as a general consumer-facing model.
- Enterprise procurement teams increasingly shift toward usage-based or short-term AI contracts rather than annual commitments, a direct response to how quickly pricing and product cycles have moved across all three labs since mid-2026.
What Developers Should Do Now That Gemini 3.8 Flash Is Confirmed
Teams evaluating Flash-tier models now have a straightforward call to make: Gemini 3.8 Flash and Gemini 3.8 Flash Cyber have joined Gemini 3.7 Flash, GPT-5.6 Luna, and Gemini 2.5 Flash-Lite on the list of models with confirmed pricing, availability, and documentation. The 1-million-token context window that circulated in leaks is now a confirmed spec rather than a claim, and the $0.75/$3.75 per million token pricing carried over from 3.7 Flash without change. The main open question left is Gemini 3.8 Flash Cyber’s long-term pricing, since Google hasn’t broken that out separately from the base model yet.
A more durable approach is still to build model-routing flexibility into new AI features, so that swapping in a cheaper or faster model — whether that’s Gemini 3.8 Flash, Gemini 3.8 Flash Cyber, another OpenAI price cut, or Anthropic’s next Sonnet revision — doesn’t require rewriting application logic later. Given how often pricing has shifted across all three labs since July, that flexibility is worth more right now than optimizing for any single model’s current numbers.
Frequently Asked Questions
Is Gemini 3.8 Flash officially released?
Yes. Google officially confirmed Gemini 3.8 Flash on September 2, 2026, alongside a security-focused variant called Gemini 3.8 Flash Cyber, in a joint launch post from Tulsee Doshi and Raluca Ada Popa. Before that, it existed only as an internal preview undergoing testing on Google’s Jetski platform under the code name “Skimaki.”
What is Gemini 3.8 Flash Cyber?
Gemini 3.8 Flash Cyber is a cybersecurity-focused variant of Gemini 3.8 Flash that Google confirmed in the same September 2, 2026 launch post, led jointly by Tulsee Doshi and security and privacy research lead Raluca Ada Popa. It’s built for tasks like vulnerability triage and security-log analysis rather than general-purpose use, and Google has not yet disclosed pricing separate from the base model.
Did Gemini 3.8 Flash launch on Wednesday, September 3, 2026, as rumored?
Not exactly. Google confirmed Gemini 3.8 Flash one day earlier, on Tuesday, September 2, 2026, ahead of the Wednesday date that had been circulating in leaks and social posts.
What was Gemini 3.8 Flash’s internal code name before launch?
Reporting before the official announcement said the project was internally code-named “Skimaki.” The earliest public trace of it was a Tencent research paper dated August 10, 2026, weeks before Google confirmed anything publicly.
How much does Gemini 3.8 Flash cost?
Gemini 3.8 Flash launched at $0.75 per million input tokens and $3.75 per million output tokens — the same rate Gemini 3.7 Flash introduced at, confirmed in Google’s September 2, 2026 launch post. Google has not broken out separate pricing for the Gemini 3.8 Flash Cyber variant.
Does Gemini 3.8 Flash really have a 1-million-token context window?
Yes. Google confirmed a 1-million-token context window for Gemini 3.8 Flash as part of its September 2, 2026 launch, matching the figure that had circulated in community posts before the official announcement.
How does Gemini’s pricing compare to GPT-5.6 and Claude 5 right now?
Gemini 3.1 Pro matches GPT-5.6 Terra at $2/$12 per million tokens. Gemini’s Flash-Lite tier at $0.075/$0.30 undercuts GPT-5.6 Luna’s post-cut $0.20/$1.20 pricing, and confirmed Gemini 3.8 Flash pricing of $0.75/$3.75 sits between the two. Claude Sonnet 5 sits at $2/$10, slightly cheaper on output than Terra and Gemini 3.1 Pro.
Why did OpenAI cut GPT-5.6 Luna and Terra pricing so sharply?
OpenAI framed the July 30, 2026 cuts as a response to businesses scrutinizing AI spend as usage scales, according to CNBC’s reporting. Luna dropped 80% and Terra dropped 20%, while flagship model Sol held its price.
Did Anthropic really cancel a planned Claude Sonnet 5 price increase?
Yes. Sonnet 5 launched at a promotional $2/$10 per million tokens, with a standard $3/$15 rate originally scheduled for September 1, 2026. Anthropic cancelled that increase and made the lower launch price permanent.
What should developers do now that Gemini 3.8 Flash is confirmed?
Build against Gemini 3.8 Flash and Gemini 3.8 Flash Cyber directly — both are generally available with confirmed pricing and a 1-million-token context window. It’s still worth designing AI features with model-routing flexibility, so the next round of price cuts or model updates from any provider can be adopted without a rebuild.
