跳到主要内容

Gemini 3.8 Flash for Claude Users: The Workhorse That Now Ties Opus 5 on DeepSWE (Sep 2026)

进阶

Google shipped Gemini 3.8 Flash on September 2, 2026, twenty days after 3.7 Flash and its third Flash release in six weeks. Same tier, same 1M window, same $0.75 / $3.75 introductory sticker, same model-ID pattern. What changed is the shape of the model's behavior: Google says 3.8 Flash "works harder," executing extra reasoning steps and calling tools iteratively on complex tasks, and the benchmark table shows what that buys. DeepSWE v1.1 moved from 65.3 to 73.7, within a third of a point of Claude Opus 5. Terminal-bench 2.1 reached 89.4, a hair above Opus 5's 89.1. And on Terminal-bench 4.0, the harder successor, 3.8 Flash posts 19.1 against Opus 5's 51.8.

This lesson is for a Claude user deciding whether to route work to it. It covers the numbers that are load-bearing, the token-usage cost of "working harder," the migration checklist that has not changed since 3.7, the gated Cyber variant, and the workloads where Claude is still the right default.

What you'll learn
  • Recognize what shipped on September 2, 2026 and why a three-week Flash cadence (3.6 → 3.7 → 3.8) is the real signal
  • Read the benchmark table with the right lens: near-parity with Opus 5 on DeepSWE and Terminal-bench 2.1, a 33-point gap on Terminal-bench 4.0, a lead on domain-agent benchmarks
  • Price the 'works harder' behavior: ~30% more output tokens per task, so the same sticker costs more per task than 3.7 Flash
  • Migrate a 3.7 Flash or Claude call to gemini-3.8-flash without surprises (the 3.6-era API removals still apply)
  • Know what Gemini 3.8 Flash Cyber is, what the Fairwind Program gates, and why you almost certainly cannot call it

The one-sentence version​

Gemini 3.8 Flash is the same-tier successor to 3.7 Flash, model ID gemini-3.8-flash, 1M-token context, 64K max output, thinking_level of low / medium (default) / high, priced at $0.75 in / $3.75 out per 1M tokens through December 31, 2026 and $1.50 / $7.50 from January 1, 2027. It is live in the Gemini API, Google AI Studio, Antigravity, Android Studio, Stitch, Gemini Enterprise, the Gemini app (AI Pro and Ultra), AI Mode in Search, and Google Sheets. Alongside it, Gemini 3.8 Flash Cyber replaces 3.5 Flash Cyber for vetted defenders through the new Fairwind Program.

The cadence is the story again. 3.6 Flash on July 21, 3.7 on August 13, 3.8 on September 2: Google is iterating its workhorse tier every three weeks while the flagship stays quiet (The Register notes 3.5 Pro missed its June date). Read that as a fast developer feedback loop at the tier most teams actually pay for. Anthropic's answer this month is Claude Fable 5.1 at the top of the stack, not a cheaper Sonnet; OpenAI's is GPT-6 Astra.

Three things about this launch that surprise people​

1. Parity with Opus 5 on two headline benchmarks, a 33-point gap on the third​

Google's evaluation report, the rows a Claude shop should care about (higher is better):

Benchmark3.8 Flash3.7 FlashClaude Opus 5Claude Sonnet 5GPT-5.6 Sol
DeepSWE v1.1 (real bug fixes)73.7%65.3%74.0%53.8%72.7%
Terminal-bench 2.189.4%—89.1%—88.8%
Terminal-bench 4.019.1%11.2%51.8%12.4%37.3%
OSWorld-2.0 (computer use)59.0%50.6%75.4%42.6%62.6%
HLE-Verified54.9%53.6%54.4%31.0%54.5%
GDPval-AA v2 (Elo)15451482182415841710
Vals Finance Agent v261.4%59.0%58.6%53.9%53.8%
Harvey Legal Agent10.0%8.8%6.7%5.0%2.5%
BioMysteryBench (hard)56.5%43.5%49.4%34.1%44.7%

Three reads. First, the DeepSWE +8.4 is the load-bearing number, exactly as the +16.3 was for 3.7 Flash: Google keeps pouring the workhorse's gains into repository-scale coding. Second, Terminal-bench 2.1 vs 4.0 is the same split we flag on DeepSeek-V4.1-Flash the following week: a cheap model can saturate the older suite while the harder one, with longer adaptive shell sessions, still separates Opus 5 by more than 30 points. Third, on the domain-agent benchmarks (finance, legal, biology) 3.8 Flash now leads every column, including Opus 5. If your work is a finance or legal agent, that row is worth an eval on its own. If your work is an interactive coding session, the 4.0 row is the one to budget around.

2. "Works harder" is a token-usage decision Google made for you​

Google's own description of the change: on complex tasks the model "exhibits greater diligence — executing extra reasoning steps, and calling tools iteratively." The launch post also says developers can "utilize lower effort levels to minimize token overhead," which is a polite way of saying the default now spends more. Emergent's read of the evaluation report puts it at roughly 30% more output tokens per task; Artificial Analysis measured $0.58 per Intelligence Index task, about 40% more than 3.7 Flash, at the same per-token sticker.

For a Claude user this is the familiar effort trade-off wearing a different name: the model reaches for more tool calls and more thinking before it answers. The sticker did not move; the bill per task did. Budget on task cost, not token price. And re-run any latency SLO you had on 3.7 Flash: more iterations means more wall-clock, even at Flash speed. The physics is on Why agents burn tokens.

3. Still cheap next to Claude, but the ratio depends on which Claude​

Artificial Analysis puts 3.8 Flash at 59 on its Intelligence Index (at high reasoning), tied with GPT-5.6 Sol at extra-high and Grok 4.6 at medium, versus Claude Opus 5 at 63 and Claude Fable 5.1 at 66 (both at max). Its cost-per-task figure for Fable 5.1 is about $3.76, roughly 6× 3.8 Flash's $0.58. Against Opus 5 on list price, Vellum's arithmetic is about 6.7× cheaper on both input and output. Against Sonnet 5 the gap is far smaller, and Sonnet 5 still leads on GDPval-AA v2 (1584 vs 1545). So the honest comparison is: 3.8 Flash is a Sonnet-tier price with DeepSWE and domain-agent scores that now reach Opus territory, and Terminal-bench 4.0 / OSWorld scores that do not. Current Claude list prices are on Models and pricing.

Model card at a glance​

SpecGemini 3.8 Flash
Model IDgemini-3.8-flash
ReleaseSeptember 2, 2026
Context window1M input tokens
Max output64K tokens
Thinking levelslow, medium (default), high; minimal not supported
Modalities intext, image, audio, video
Modalities outtext
Knowledge cutoffMarch 2026 (some domains limited to January 2025)
Removed parameterstemperature, top_p, top_k, candidate_count (since the 3.6 line)
Intro price (through Dec 31, 2026)$0.75 in / $3.75 out per 1M tokens
Standard price (from Jan 1, 2027)$1.50 in / $7.50 out per 1M tokens
AvailabilityGemini API, Google AI Studio, Antigravity, Android Studio, Stitch, Gemini Enterprise, Gemini app (AI Pro/Ultra), AI Mode, Google Sheets

The knowledge cutoff line is new information versus the 3.7 page and worth a second look: March 2026 headline, but the model card says some domains are limited to January 2025. If you rely on the model knowing a 2026 library API, verify with a grounding tool or paste the docs. That is the same advice we give for any model in Freshness half-life.

Migrating from 3.7 Flash (or from Claude)​

The API surface is unchanged from 3.7 Flash. The list of things that break a Claude port is also unchanged, and it keeps catching people.

Guided walkthrough1 of 7
  1. Use gemini-3.8-flash. The 3.7 endpoint keeps working, but the benchmark deltas and the 'works harder' behavior only arrive with the new ID.

Minimal 3.8 Flash call (Python SDK)

from google import genai

client = genai.Client()  # reads GOOGLE_API_KEY

resp = client.models.generate_content(
  model="gemini-3.8-flash",
  contents="Review this diff. List the 3 riskiest hunks as file:line with a one-sentence reason each.",
  config={
      "thinking_level": "low",   # start low on 3.8 Flash; it already 'works harder' at medium
      # no temperature / top_p / top_k / candidate_count
      "tools": [{"function_declarations": [...]}],
  },
)
print(resp.text)

Gemini 3.8 Flash Cyber and the Fairwind Program​

Alongside the general model, Google released Gemini 3.8 Flash Cyber, replacing 3.5 Flash Cyber. Google's numbers: CWE-Bench patching at 47.2% pass@1 against "a leading frontier model" at 47.8%; a vulnerability-discovery success rate above 70% across 20 programming languages; 2.6× more correct patches to Chrome vulnerabilities than the best commercial models per Chrome Security; and, per Wiz, 7.5–9.7 points higher recall at 2.3–5.2× lower cost.

You almost certainly cannot call it. Access runs through the Fairwind Program, limited to "trusted government authorities, as well as critical infrastructure operators and software maintainers," by application. This is the same shape as OpenAI gating GPT-6 Astra's exploit capability behind a vetted program (see the Astra field guide) and Anthropic's treatment of Mythos-class models. For a Claude user the practical takeaway is that the frontier's cyber capability now ships as a separate, gated SKU on all three stacks, and the general models are deliberately not it. Our security section covers what that means for the agents you can call.

When to reach for 3.8 Flash, and when to stay on Claude​

Pro tip
  • Reach for 3.8 Flash: repository-scale coding that looks like DeepSWE, where 73.7 vs Opus 5's 74.0 at a sixth of the task cost is a real trade.
  • Reach for 3.8 Flash: finance, legal, and life-science agent workflows — it now leads every column on Vals Finance Agent v2, Harvey Legal, and BioMysteryBench, including Opus 5.
  • Reach for 3.8 Flash: long-video and chart understanding (LVBench 87.8, CharXiv 86.2) where Claude has no equivalent input path.
  • Stay on Claude Opus 5: interactive terminal agents and computer use. Terminal-bench 4.0 (51.8 vs 19.1) and OSWorld-2.0 (75.4 vs 59.0) are the widest gaps on the table.
  • Stay on Claude for knowledge work where GDPval-AA v2 matters: Opus 5 at 1824 and even Sonnet 5 at 1584 are ahead of 3.8 Flash's 1545.
  • Stay on Claude for anything built on prefilled turns, temperature control, or Anthropic's tool-use loop — you will fight the Gemini API rather than use it.
  • Cross-check both, at the effort you will ship: 3.8 Flash's table is at default or high thinking and costs ~30% more tokens per task than 3.7; Claude's is at max effort. Neither is your production setting.

For the broader framing, Gemini for Claude users; for the coding-specific comparison, Claude vs GPT vs Gemini for coding; for the durable decision framework, Choosing a model.

Quick check​

Check yourself

0/5
  1. Gemini 3.8 Flash scores 89.4 on Terminal-bench 2.1 (Opus 5: 89.1) and 19.1 on Terminal-bench 4.0 (Opus 5: 51.8). What is the right takeaway?
  2. Same $0.75 / $3.75 sticker as 3.7 Flash. Why can a task cost more on 3.8 Flash?
  3. Which of these will break a Claude call ported to gemini-3.8-flash?
  4. Who can call Gemini 3.8 Flash Cyber?
  5. On which benchmark family does 3.8 Flash lead every column in Google's table, including Opus 5?
按 Enter 或空格键翻转卡片。使用左右方向键在卡片之间切换。已显示术语。
1 / 8
Key takeaways
  • DeepSWE +8.4 to 73.7 is the load-bearing number: a Sonnet-priced model now within 0.3 of Opus 5 on repository-scale bug fixing.
  • Terminal-bench 4.0 (19.1 vs 51.8) and OSWorld-2.0 (59.0 vs 75.4) are where Opus 5 stays far ahead — interactive agents and computer use stay on Claude.
  • 'Works harder' means ~30% more tokens per task at the same sticker. Budget per task, start thinking_level at low, and re-check latency SLOs.
  • The 3.6-era API removals (temperature/top_p/top_k/candidate_count, no prefill, call_id + name on FunctionResponse) still break Claude ports; the migration list has not shrunk.
  • 3.8 Flash Cyber is a separate, Fairwind-gated SKU — the third stack this month to ship frontier cyber capability as something you cannot simply call.

Sources & further reading​

Next​