On September 2, 2026 (US Pacific time), Google DeepMind released Gemini 3.8 Flash and Gemini 3.8 Flash Cyber together, with Chinese coverage on September 3. This is the third Flash-tier model in roughly six weeks, after 3.6 Flash on July 21 and 3.7 Flash on August 13. Google positions 3.8 Flash as the "most intelligent Flash," built for long-horizon software engineering, autonomous agents, and enterprise workflows. The interesting part is not that another model arrived. It is that Google iterates this cheap, high-throughput line at a rare pace while putting a cost truth on the table that many miss: list prices did not move and the model got smarter, yet because it is allowed to "work harder," the actual cost per task goes up. This article labels every number by source so you can quote it honestly.
1. Release context: the third Flash in six weeks
Start with the cadence. From Gemini 3.6 Flash on July 21 to 3.7 Flash on August 13 to 3.8 Flash on September 2, Google shipped three Flash-tier models in six weeks. That speed is unusual among the major labs. Flagships typically land quarterly, and the Flash tier used to ride along with the flagship rather than rolling on its own fast clock. Stating the cadence makes one point: Google now runs the Flash tier as an independent, rapidly polished line whose purpose is to keep feeding capability gains into the cheap half of the catalog.
The practical implication is that evaluation conclusions expire faster. The judgment you formed about 3.7 Flash last month may be rewritten by 3.8 Flash this month. Every cross-generation comparison here uses 3.7 Flash as the nearest baseline, and whenever version comparison appears, keep that three-in-six-weeks timeline in mind. Note too that both models shipped together; the security-specialized sibling takes a trusted-access route, so the two lines diverge in availability from the start.
The other piece of context is Gemini 3.8 Flash Cyber, announced in the same breath. It is not a simple variant of 3.8 Flash but a security-specialized sibling line, distributed only to vetted defenders through the new Fairwind Program, with no public price. More on Cyber in section five; for now put it on the same release table as the main model to remind readers that what shipped and what you can obtain are two different things.
2. Specifications: million-token context and three thinking levels
The core specs of 3.8 Flash are laid out below.
| Spec | Gemini 3.8 Flash | Comparison note |
|---|---|---|
| Input context | 1,048,576 tokens | 1M, sustaining the Flash-tier million-token context |
| Max output | 65,536 tokens (64K) | Text output cap; note 64K, not the flagship 128K |
| Input modalities | text, image, video, audio, PDF | Multimodal input |
| Output modalities | text only | No multimodal generation in output |
| Thinking levels | low / medium / high | User selects reasoning intensity per task |
| Knowledge cutoff | March 2026 (some domains Jan 2025) | Some domains cut off earlier; watch this |
A few spec points deserve attention. First, the output cap is 64K rather than the flagship 128K. That is enough for most agent pipelines, but leave headroom for very long generations. Second, thinking comes in three levels rather than "adaptive always on," which hands engineering a clear lever: low for simple tasks to save tokens, high for hard tasks to trade quality. That lever is exactly the switch behind the per-task cost discussion. Third, the knowledge cutoff sits in January 2025 for some domains, so when citing recent facts there, bring retrieval or tools rather than assuming the model knows.
Availability is broad: Gemini API, Google AI Studio, Android Studio, Antigravity, Gemini Enterprise, the Gemini app, and AI Mode in Search. For teams this means you can probably use it without changing platforms, but test capability consistency across entry points, because different surfaces may wrap different guardrails and default parameters.
3. Benchmarks: coding is the strength, OS and long-horizon are the gaps
The methodology caveat first: the Google-owned benchmarks below (DeepSWE, Terminal-Bench, SWE-Bench Pro, HLE-Verified, Vals, Harvey) are Google's published numbers, not independent third-party measurement. A few rows such as Terminal-Bench 4.0 come from third-party or later reporting and are flagged separately. We have not seen an independent organization replicate these under like conditions, so read for magnitude and direction, not as settled values.
| Benchmark | Gemini 3.8 Flash | Published comparison | Source note |
|---|---|---|---|
| DeepSWE v1.1 | 73.7% | Claude Opus 5 74.0%, GPT-5.6 Sol 72.7%, 3.7 Flash 65.3% | Google; near Opus 5 |
| Terminal-Bench 2.1 | 89.4% | Opus 5 89.1% | Google; labeled world #1 |
| SWE-Bench Pro | 61.6% | no comparison given | |
| HLE-Verified | 54.9% | no comparison given | |
| Terminal-Bench 4.0 | 19.1% | Opus 5 51.8%, Fable 5.1 55.8%, GPT-5.6 Sol 37.3% | clear weakness |
| OSWorld 2.0 | 59.0% | Opus 5 75.4% | clear weakness |
| Vals Finance Agent v2 | 61.4% | — | Google; labeled #1 |
| Harvey Legal Agent | 10.0% | — | Google; labeled #1 |
Read the table in two layers. The strength layer: DeepSWE 73.7% sits only 0.3 points below Opus 5, and Terminal-Bench 2.1 at 89.4% is labeled world number one yet only 0.3 points above Opus 5 at 89.1%, so they are in the same band. These two support the claim that frontier-class coding has entered the cheap tier. Worth flagging: 89.4% differs from the 90.8% some outlets report, and this article uses Google's figure.
The weakness layer matters more. Terminal-Bench 4.0 is only 19.1%, less than half of Opus 5 at 51.8% and clearly below Fable 5.1 at 55.8% and GPT-5.6 Sol at 37.3%. OSWorld 2.0 at 59.0% trails Opus 5 at 75.4% by about sixteen points. Those gaps show 3.8 Flash is still clearly weaker than the flagship on tasks demanding very long horizons, multi-step tool chains, and stable GUI operation. Its coding strength concentrates in verifiable closed loops like writing code and running a terminal, while a real gap remains on long autonomous operation of real systems. Do not extrapolate the high DeepSWE score across an entire workflow.
The two vertical agent benchmarks took labeled number-one spots (Vals 61.4%, Harvey 10.0%), but both lack a same-condition comparison and Harvey is low in absolute terms, so quote them with the qualifier "labeled, no control group" and do not treat them as reproducible commitments. For a cross-vendor comparison of flagship and Flash-tier coding and reasoning, see the flagship coding and reasoning review. For a lighter vision-experiment approach to building evaluations close to your workload, see the DeepSeek V4 Flash vision experiment resource.
4. Cost: list price flat, per-task price up
On pricing, Google offered a same-price-for-a-limited-time arrangement, so the schedule is laid out first.
| Line item | Introductory (through 2026-12-31) | Standard (from 2027-01-01) |
|---|---|---|
| Input | $0.75 per million tokens | $1.50 per million tokens |
| Output | $3.75 per million tokens | $7.50 per million tokens |
The introductory price is identical to 3.7 Flash, so no increase. From January 1, 2027 the standard price returns at exactly double. Therefore "same price, smarter" holds at the list-price level, provided you migrate workloads inside the introductory window. That window itself is an action in the checklist; do not remember it only at the end of December.
The real story is per-task cost. The key mechanism: Google lets the model "work harder" with more reasoning steps, more iterative tool calls, and longer agentic loops. Artificial Analysis measured single-task cost rising about 40%, from $0.40 to $0.58. In other words, the "cheaper" on the price sheet is partly eaten by a longer work path; the model gets smarter by spending more tokens, a third-party figure not Google's, so cite it as such.
This nuance heavily affects the ledger. If your workload is short and flat, list price did not move and the model got smarter, so you win. But if the task is a long-horizon agent that retries and loops through tools, the bill scales with token inflation rather than with one call. Against Opus 5, 3.8 Flash's unit price is roughly one seventh, but that premise holds only if you control actual token consumption per task rather than letting the model run long loops throughout.
How do you contain that inflation? Two routes. First, tune the thinking level down per task in code, using low for low-risk batch work to save tokens. Second, put reusable context into cache to reduce repeated reads. The latter is exactly the logic in the agentic cache cost comparison review: the token cost in an agent pipeline is dominated by re-reading the same system prompt, tool definitions, and long history. In short, 3.8 Flash's cost-performance is not automatic; it depends on whether you govern the "work harder" switch.
5. Gemini 3.8 Flash Cyber: a new distribution model for security models
Gemini 3.8 Flash Cyber, announced alongside, lives at the other end of the track. It is a security-specialized sibling, but its distribution is entirely different from the main model: it goes only to vetted defenders through the new Fairwind Program on trusted access, with no public price. This article cannot confirm its pricing and marks it unconfirmed.
Cyber's existence sends two signals. First, Google states explicitly that security-direction training also lifted the base model's coding, meaning part of 3.8 Flash's coding progress comes from spillover of security training. That explains how a Flash tier that "writes better code" gains benefit from a security line never publicly released; the two lines feed each other in capability. Second, and more worth remembering: a security-specialized model no longer takes the old path of "public release plus uniform price" but instead the gated path of "vetted defenders plus trusted access."
This mirrors Anthropic limiting Mythos 5.1 to US institutions, signaling the major labs are formally decoupling "how strong" from "who may obtain." For engineering teams the direct implication is that Cyber is probably not a console one-click option. Its value is closer to a capability-ceiling announcement plus a distribution-model demo. What lands in your workflow is the spillover benefit to coding and security tasks in the main 3.8 Flash, not Cyber itself. Keep the two separate: the former is immediately evaluable, the latter absent until the Fairwind Program opens to your institution.
6. What to do now
| Action | Rationale | Priority |
|---|---|---|
| Add 3.8 Flash to coding and agent evaluation | DeepSWE 73.7%, Terminal-Bench 2.1 89.4%, near flagship | High |
| Tune thinking level per task to govern tokens | Per-task cost up ~40% from working harder (third-party) | High |
| Do not fully switch long-horizon OS / multi-step GUI tasks | Terminal-Bench 4.0 19.1%, OSWorld 2.0 59.0% clearly weak | High |
| Catch the introductory price window | Same price through 2026-12-31, doubles in 2027 | Medium |
| Use cache architecture to contain per-task cost | See cache cost comparison | Medium |
| Treat Cyber as a capability note, not an option | Fairwind trusted access only, price unconfirmed | Medium |
Rows two, three, and five connect. First use row three to identify tasks 3.8 Flash still does poorly (do not hard-switch, do not be misled by DeepSWE), then use row two to dial tokens down on tasks it does well by thinking level, and finally use row five's cache architecture to pull the per-task bill back to the one-seventh of Opus range. These three steps are the most concrete landing order here, and the only way to turn "smarter but pricier" into "smarter and cheaper."
The part worth remembering is not the coding score near Opus, but three stacked facts: the cheap tier rolling at a rare cadence, per-task cost inflating because the model works harder, and security models moving to gated distribution. It reminds engineering that cost-performance is not a factory setting but the result of governing how hard the model works, and it reminds procurement that the cheap tier's ceiling climbs fast toward the flagship while availability is re-sliced by risk.
References
- Google DeepMind official blog and release notes (Gemini 3.8 Flash / 3.8 Flash Cyber, 2026-09-02): "most intelligent Flash" positioning; six-week three-ship cadence with 3.6 Flash (Jul 21) and 3.7 Flash (Aug 13); million-token input context; 64K output; multimodal input and text output; three thinking levels; March 2026 cutoff (some domains Jan 2025); availability on Gemini API, AI Studio, Android Studio, Antigravity, Gemini Enterprise, Gemini app, Search AI Mode; introductory $0.75 / $3.75 and 2027 standard $1.50 / $7.50 pricing; Google-owned benchmarks (DeepSWE 73.7%, Terminal-Bench 2.1 89.4%, SWE-Bench Pro 61.6%, HLE-Verified 54.9%, Vals 61.4%, Harvey 10.0%); statement that security training lifted base-model coding; 3.8 Flash Cyber with Fairwind Program trusted access. All except Cyber pricing are Google's published figures; URL unconfirmed.
- Artificial Analysis (third-party): single-task cost rising about 40% from $0.40 to $0.58 from higher reasoning effort. Third-party measurement, not Google's line, and varies with task and effort.
- Tencent News and ITHome (2026-09-03): Chinese cross-check of timing and details; some outlets report Terminal-Bench 2.1 at 90.8%, a spread from Google's 89.4%, resolved here in favor of the Google figure.
- 9to5Google, VentureBeat, MarkTechPost: English reporting cross-checking launch highlights, Flash-tier cadence, and Cyber gated distribution.
- Comparison baselines (control points in Google's figures): Claude Opus 5 (DeepSWE 74.0%, Terminal-Bench 2.1 89.1%, Terminal-Bench 4.0 51.8%, OSWorld 2.0 75.4%), GPT-5.6 Sol (DeepSWE 72.7%, Terminal-Bench 4.0 37.3%), Claude Fable 5.1 (Terminal-Bench 4.0 55.8%) as cross-model references.