A Hacker News discussion around Gemini 3.8 Flash had reached 575 points and 355 comments by 16:30 UTC on launch day. The number developers will still care about in January sits in Google's pricing footnote: the new model costs $0.75 per million input tokens and $3.75 per million output tokens through December 31, then both rates double. Google's API guide also says the model may consume more tokens on difficult jobs because it takes smaller reasoning steps, calls tools repeatedly, and checks its work. A production bill can therefore rise through the rate and the number of tokens used.
Google released Gemini 3.8 Flash on September 2 alongside Gemini 3.8 Flash Cyber. This is the company's third Flash release in six weeks, and it arrives only three weeks after 3.7 Flash, according to the launch announcement. The general model is available through the Gemini API and several Google products. The cyber version has looser security mitigations and is limited to approved defenders through a new access program. Those products share a foundation, but their distribution, safeguards, and likely buyers are quite different.
The introductory rate hides the January bill
The launch rate matches 3.7 Flash, yet it lasts for four months. One million input tokens plus one million output tokens cost $4.50 at the introductory rate and $9 from January 1, based on the published API prices. A sample monthly workload using 100 million input tokens and 20 million output tokens would move from $150 to $300 before caching, tool, or other service charges. That calculation assumes token use stays flat. Google has not published a typical token-use multiplier for 3.8 versus 3.7.
Token use is partly under the developer's control. Gemini 3.8 Flash supports low, medium, and high thinking levels, with medium as the default; sending minimal returns an error. Google assigns low to latency-sensitive work and high to difficult reasoning or tool orchestration in its developer documentation. A Python request can set the lower level explicitly:
generation_config={
"thinking_level": "low"
}
That switch should be measured on the application rather than treated as a billing toggle. Google says lower effort reduces token use, while medium and high spend more work on difficult requests. For an agent that repairs code or operates software, a shorter answer that fails may cost more after retries than a longer answer that finishes once. The useful comparison is the cost of an accepted result, including tool calls and failed loops, under each setting described in the 3.8 Flash guide.
The benchmark chart mixes several measuring methods
Google's headline result for broad reasoning is 54.9 percent on HLE-Verified. It also says 3.8 Flash beats most larger models on DeepSWE v1.1 and improves on 3.7 Flash in finance and legal agent tests. These numbers come from the company's launch post, which presents the new model as a lower-cost option for long-running coding and knowledge work.
The accompanying evaluation methodology puts boundaries around that comparison. Google computed its own DeepSWE result with a mini-swe agent harness and high thinking, while it took other models' scores from the public leaderboard. Non-Gemini figures generally come from providers' self-reported results. For some computer-use tests, Google maximized results over three runs and added batched tool calls to its harness. The chart collects results produced through several routes rather than one independent laboratory run under identical conditions.
HLE-Verified is more tightly described. Google tested the full 1,811-item verified set and excluded 689 questions from the original Humanity's Last Exam set that researchers had marked uncertain. Its methodology also notes that content filters blocked a substantial share of questions for Sonnet 5 and a small number for Opus 5. Those details do not erase the 54.9 percent result, but they matter when comparing columns or trying to reproduce the test from the evaluation document.
Teams considering a migration need their own fixed task set. Record whether the model completes each job, how long it takes, how many tokens it spends, and which tool actions require human correction. That test can reveal whether a higher benchmark score survives contact with a particular repository or workflow. It also makes the January price change part of the decision instead of a surprise after integration.
The model card is frank about several limits
Gemini 3.8 Flash accepts text, images, audio, and video, with a context window up to one million tokens and a maximum text output of 64,000 tokens. Its public model card lists hallucinations, occasional slowness or timeouts, and extra token use at higher effort among the known limits. The knowledge cutoff is March 2026 for some material, while knowledge in other domains may stop at January 2025.
The same card reports a slight regression in multilingual safety against 3.7 Flash. The automated multilingual measure rose 5.4 percentage points, and lower is better on that test. Text-to-text safety improved by 0.4 points, image-to-text safety was unchanged, and unjustified refusals worsened by 1.1 points. Google says a manual review found that the recorded losses were overwhelmingly false positives or did not involve severe material. The company also warns that its updated evaluation sets make some results unsuitable for direct comparison with older model cards.
Google did not run a fresh full frontier assessment for the general 3.8 Flash model. The model card says the company found no material capability increase over 3.7 Flash in the domains tracked by its Frontier Safety Framework, so it relied on the prior model's assessment and judged 3.8 unlikely to reach a tracked or critical capability level. That conclusion applies to the public Flash model. Flash Cyber is presented separately and receives different mitigations and access rules.
Cyber capability comes with access rules
Gemini 3.8 Flash Cyber is available through Google's new Fairwind Program, which combines the model with the CodeMender system for finding, checking, and repairing vulnerabilities. Google says the program has more than 650 partners across government, critical infrastructure, and technology. Participating organizations must restrict access to internal security, incident response, or penetration-testing staff and require protections such as multi-factor authentication.
Most published results for the cyber model still need outside replication. Google reports a success rate above 70 percent on an internal vulnerability-discovery test spanning 20 programming languages. On the external CWE-Bench patching test, it reports 47.2 percent pass@1, close to a leading larger model at 47.8 percent. The launch post also says Chrome's security team obtained 2.6 times as many correct patches as it did from the best larger commercial models it tested.
Two more figures come with the same attribution limits. Wiz reported 7.5 to 9.7 percentage points more recall on its internal penetration-testing benchmark at 2.3 to 5.2 times lower cost than other leading models. Google's Cloud Vulnerability Research team says Flash Cyber found a critical foundational vulnerability in under two hours. The announcement does not identify the flaw, the comparison models, or enough test detail for outsiders to reproduce either result.
Access control is part of the product design. The general Flash model ships with stricter controls against cyber offense, while Flash Cyber uses more permissive mitigations so approved defenders can inspect and repair dangerous code. Google says Fairwind gives defenders an early period to harden systems before similar capabilities spread, and it plans to expand access with input from governments, companies, and open-weight communities. No public price for Flash Cyber appears in the launch material.
Migration involves more than changing the model name
The production model ID is gemini-3.8-flash, and Google labels it generally available. Its migration checklist tells developers to remove temperature, top_p, and top_k, replace thinking_budget with thinking_level, and stop using candidate_count. Multi-turn applications should use the server-side previous_interaction_id; applications using the older generateContent API also need to preserve thought signatures and supply identifiers on function responses.
Google says 3.7 Flash remains supported, and the 3.8 announcement gives no retirement date. That leaves room to run both models against saved production tasks before switching. It is especially useful for agent systems, because 3.8 is now the default model in Google's Antigravity managed agent and can take more iterative tool steps than its predecessor, according to the API guide.
The next four months should produce evidence that the launch charts cannot: token distributions at each thinking level, independently repeated coding scores, and failure rates in long-running agents. For Flash Cyber, watch for disclosed vulnerabilities and patches that outside maintainers can inspect. The hard date is already known. Google.s published schedule doubles the public model.s token rates on January 1, so teams evaluating 3.8 now need to test against that price rather than the temporary one.