On September 2, 2026, Google announced "Gemini 3.8 Flash" and its security-focused counterpart, "Gemini 3.8 Flash Cyber." This marks the third update in six weeks—three weeks after 3.7 Flash and six weeks after 3.6 Flash. The standard version is generally available at the same time-limited pricing as its predecessor, while the Cyber version is offered only to vetted defensive-side organizations that pass a review process. Google's own evaluation tables show substantial improvements, but once you factor in Arena's human preference ratings, it's still too early to say this model has uniformly replaced the top-tier models.
Up to 73.7% in three weeks—where gains appeared, and where gaps remain
In Google's own apples-to-apples comparison table, 3.8 Flash outperformed 3.7 Flash on several benchmarks: 73.7% vs. 65.3% on DeepSWE v1.1, 61.4% vs. 59.0% on the Finance Agent benchmark, 10.0% vs. 8.8% on the Legal Agent benchmark, and 54.9% vs. 53.6% on HLE-Verified. On the other hand, it scored only 19.1% on Terminal-bench 4.0, well behind Claude Opus 5's 51.8% and GPT-5.6 Sol's 37.3%, and 59.0% on OSWorld-2.0, below Opus 5's 75.4%.
- Gemini 3.8 Flash
- Gemini 3.7 Flash
- Claude Opus 5
- GPT-5.6 Sol
データを表で見る
| Gemini 3.8 Flash (%) | Gemini 3.7 Flash (%) | Claude Opus 5 (%) | GPT-5.6 Sol (%) | |
|---|---|---|---|---|
| DeepSWE v1.1 | 73.7 | 65.3 | 74 | 72.7 |
| Finance Agent v2 | 61.4 | 59 | 58.6 | 53.8 |
| Legal Agent | 10 | 8.8 | 6.7 | 2.5 |
| Terminal-bench 4.0 | 19.1 | 11.2 | 51.8 | 37.3 |
| HLE-Verified | 54.9 | 53.6 | 54.4 | 54.5 |
| OSWorld-2.0 | 59 | 50.6 | 75.4 | 62.6 |
The biggest gains cluster around multi-step software development tasks carried through to completion, financial and legal workflow processing, and understanding of charts and long-form video. On BioMysteryBench's hardest question set, scores rose from 43.5% to 56.5%, and LABBench2 climbed from 82.1% to 86.2%. By contrast, on Terminal-bench 4.0—a benchmark for general agentic capability—the model improved by 7.9 points over its predecessor but still trails Opus 5 by 32.7 points. Even when benchmark names sound similar, terminal operations, long-horizon development, and OS operation are distinct capabilities, and a single figure like 73.7% cannot be used to claim overall superiority in software development work.
The comparison table itself comes with caveats. Google places its own Gemini figures alongside self-reported or publicly disclosed figures from other companies in a single table, and there's no guarantee that every model was re-run under identical conditions by the same organization using the same API settings. Even the 54.9% on HLE-Verified is only 0.4 points ahead of GPT-5.6 Sol's 54.5% and 0.5 points ahead of Opus 5's 54.4%. Rather than focusing on rank order, it's more important to first check which task set and reasoning configuration produced each number.
Arena's 1494 is provisional—a 3-point gap isn't a verdict
Arena has users choose their preferred response from two anonymized answers, providing a signal closer to real-world usage than fixed-question accuracy scores. The "rank spread" that Arena reports alongside each score accounts for overlapping confidence intervals among models and indicates the range of ranks a model could plausibly occupy. The raw rank is simply the current best estimate—it doesn't settle superiority over a rival whose rank spread overlaps with it.
The figures used here come from an aggregation with the default style-control adjustment applied. In the Text Arena overall leaderboard for September 2, 2026, Gemini 3.8 Flash (High) held a provisional 7th place with a score of 1494±9, a rank spread of 2nd to 21st, and 5,137 votes. Gemini 3.7 Flash (High) sat at 10th place with 1491±8—a gap of just 3 points, with overlapping confidence intervals, meaning the current Arena data cannot support a clear win for 3.8.
| Text Arena Overall (Sept 2) | Score | Raw Rank | Rank Spread | Votes |
|---|---|---|---|---|
| Claude Fable 5 | 1508±5 | 1st | 1st–4th | 27,013 |
| Claude Opus 4.6 High | 1505±4 | 2nd | 1st–5th | 72,114 |
| Gemini 3.8 Flash High | 1494±9 | 7th | 2nd–21st | 5,137 |
| Gemini 3.7 Flash High | 1491±8 | 10th | 3rd–23rd | 5,682 |
There's a reason for the "provisional" label. Arena displays models that were tested anonymously before their public release as provisional until enough new votes accumulate after the public launch. The listing threshold is a minimum of 1,000 votes, but for models that carry over pre-release votes, Arena re-checks whether the evaluation has stabilized under post-launch usage conditions. 3.8 Flash already has 5,137 votes, yet its rank spread remains wide, from 2nd to 21st. The clear generational improvement shown in Google's fixed benchmarks and the small, uncertain gap shown by Arena aren't contradictory—they're simply measuring different things.
Same $0.75, but reasoning volume drives total cost
3.8 Flash's Standard pricing, in effect through December 31, 2026, is $0.75 per million input tokens and $3.75 per million output tokens (including thinking tokens)—identical to 3.7 Flash. On January 1, 2027, these rates rise to $1.50 and $7.50 respectively. Under a simple calculation using one million tokens each of input and output, the total cost doubles from $4.50 to $9.00.
- 入力
- 出力
データを表で見る
| 入力 (ドル/100万トークン) | 出力 (ドル/100万トークン) | |
|---|---|---|
| Gemini 3.8 Flash | 0.75 | 3.75 |
| Claude Opus 5 | 5 | 25 |
| Claude Sonnet 5 | 2 | 10 |
| GPT-5.6 Sol | 4 | 20 |
| GPT-5.6 Terra | 2 | 12 |
For identical processing volumes, 3.8 Flash's pricing through the end of the year is lower than Opus 5's or Sonnet 5's. But Google itself notes that for complex tasks, 3.8 Flash may take additional reasoning steps, call tools repeatedly, and consume more tokens under higher reasoning effort. Switching from 3.7 to 3.8 at the same per-token rate doesn't guarantee the same bill if output length, retries, and tool use all increase. What matters when adopting the model isn't the price per million tokens, but the total tokens, latency, and retry count needed to complete the same job.
The API's limits are 1,048,576 tokens for input and 65,536 tokens for output, and the model name is the Stable release gemini-3.8-flash. It accepts text, image, and video input, and also supports audio and PDFs. It supports code execution and function calling. Google Search and file search are available, as is structured output. Computer use is in Preview, while Live API, image generation, and audio generation are not supported. The model card lists hallucination and occasional latency or timeouts as known limitations, and its knowledge cutoff varies by domain—either March 2026 or January 2025.
The Cyber version's 86.2% isn't the general API's performance
Gemini 3.8 Flash Cyber recorded an 86.2% pass@1 (first-attempt success rate) on CyberGym, beating 3.5 Flash Cyber's 77.5% by 8.7 points and GPT-5.6 Sol's 83.6% by 2.6 points. On CWE-Bench, it scored 47.2%, just 0.6 points behind Fable 5's 47.8%, though Google positions it as Pareto-optimal when combined with cost.
データを表で見る
| pass@1 (%) | |
|---|---|
| Gemini 3.8 Flash Cyber | 86.2 |
| Gemini 3.5 Flash Cyber | 77.5 |
| GPT-5.6 Sol | 83.6 |
| Mythos 5 | 83.8 |
| GPT-5.5-Cyber | 85.6 |
3.8 Flash Cyber posts the top score in this comparison, but GPT-5.5-Cyber comes close. CyberGym's rankings cannot be directly transferred to CWE-Bench, which measures remediation capability, or to real-world success rates.
CyberGym is a benchmark for discovering vulnerabilities in C/C++ code, while CWE-Bench measures the ability to fix them. Google's internal evaluation spanning 20 languages reportedly achieved a success rate above 70%, and in Chrome Security testing, the model reportedly generated correct fixes at 2.6 times the rate of large commercial models. Wiz also reported that in its internal penetration testing, recall was 7.5 to 9.7 points higher while costs were 2.3 to 5.2 times lower. However, the task sets and comparison baselines behind these figures haven't been disclosed. Without absolute values or raw data, these claims can't be read as independently reproduced results.
Access conditions are stricter than for the standard version. The Fairwind Program claims over 650 participating organizations, but restricts access to internal security, incident response, and penetration testing personnel, and requires phishing-resistant multi-factor authentication and access tracking. Sharing, redistributing, or reselling access to the model is prohibited. No public pricing table for the Cyber version exists, and the standard version's $0.75 input / $3.75 output rates cannot simply be applied to it. The 86.2% figure reflects a restricted model run under restricted conditions—it is not the cybersecurity performance of the general-access Gemini API version of 3.8 Flash.
Why Terminal-bench figures diverge across official materials remains unexplained
On Terminal-bench 2.1, Google DeepMind's public comparison table lists 89.4% for 3.8 Flash and 85.8% for 3.7 Flash, while the Gemini Enterprise Agent Platform developer guide lists 90.8% and 81.6% for the same models. Both pages report different figures under the same benchmark name, and since neither publicly explains the difference in configuration or update timing, the generational improvement cannot be pinned to a single number.
The discrepancy is not a minor rounding error. Calculated from the first set of figures, the improvement is 3.6 points; from the second, it's 9.2 points. There's no basis for declaring either set incorrect, and it would be inappropriate to combine the higher 3.8 figure with the lower 3.7 figure to maximize the apparent gap. Until Google updates its methodology or tables, readers should only compare figures within the same source document.
3.8 Flash has, by fixed benchmarks, narrowed the gap between the Flash price tier and top-tier models. But Arena's provisional standing, total costs driven by reasoning volume, access restrictions on the Cyber version, and inconsistencies in official figures don't disappear from the leaderboard released on launch day. Teams making adoption decisions should run 3.7 and 3.8 side by side—on the same tasks, the same reasoning settings, and the same harness—during the period when pricing remains equal, through the end of the year. By logging completion rates and total token counts, along with latency and retry counts, teams can judge for themselves, using their own numbers, whether the benefits of migration hold up after the price doubles in January 2027.