How much does Gemini 3.8 Flash cost?

A promotional price of $0.75 per million input tokens and $3.75 per million output tokens runs through December 31, 2026. After January 1, 2027 prices double to $1.50 input / $7.50 output. Batch pricing is half: $0.375 / $1.875 per million. Output pricing includes thinking tokens.

Is Gemini 3.8 Flash better than Gemini 3.7 Flash?

Reviewers describe it as the first Flash-line model to meaningfully advance intelligence rather than just price or latency. Google reports DeepSWE 65.3% to 73.7%, OSWorld-2.0 50.6% to 59.0%, and Terminal-Bench 2.1 85.8% to 89.4%; Artificial Analysis puts its Intelligence Index at 59 with high reasoning, up from 56 for 3.7 Flash. The trade-off: it reasons harder, so per-task token costs rise about 30-45%.

What is the context window of Gemini 3.8 Flash?

1,048,576 (1M) input tokens and 65,536 (~64K) output tokens. It accepts text, images, video (up to 3 hours low-res or 1 hour high-res), audio (up to 9.5 hours), and PDFs, and outputs text. Output speed is roughly 300-350 tokens per second, among the fastest models tracked by Artificial Analysis.

How does Gemini 3.8 Flash compare to Claude Opus 5 or GPT-5.6 Sol?

On Artificial Analysis it ties GPT-5.6 Sol and Grok 4.6 at Intelligence Index 59 (behind Claude Opus 5's 63). Google reports it beats larger flagships on several specialist agent benchmarks (finance, legal, HLE-Verified) at roughly 1/15th the per-token sticker price, but it trails frontier flagships on long-horizon autonomous work such as Terminal-Bench 4.0 (19.1% vs Opus 5's 51.8%) and OSWorld-2.0. Figures are launch-day, vendor-reported.

Gemini 3.8 Flash Review 2026: Google's Agentic Workhorse Beats Flagships on Specialist Evals at 1/15th the Price

Q: When was Gemini 3.8 Flash released?

Google DeepMind released Gemini 3.8 Flash on September 2, 2026, the third Flash-tier update in about six weeks (3.6 Flash on July 21, 3.7 Flash on August 13). It launched on the Gemini API, Google AI Studio, Gemini Enterprise, and Google AI Pro/Ultra subscriptions, alongside the gated security twin Gemini 3.8 Flash Cyber.

Q: What is Gemini 3.8 Flash Cyber?

A cybersecurity-specialized variant of the same model with more permissive cyber mitigations, built for autonomous vulnerability discovery and patching. Google's Chrome Security team says it produced 2.6x more correct patches on real Chrome vulnerabilities than larger frontier models. It is not public: access is gated through Google's Fairwind Program for vetted defenders (governments, critical-infrastructure operators, software maintainers, named security vendors).

Gemini 3.8 Flash Review 2026: Google's Agentic Workhorse Beats Flagships on Specialist Evals at 1/15th the Price — AA 59, DeepSWE 73.7%, $0.75/$3.75

September 6, 2026 · Updated 2026 · AI Agents · AI Coding · Read time: ~9 min

Verdict first: Gemini 3.8 Flash (September 2, 2026) is the first model in Google's budget "Flash" line that does not just get cheaper or faster — it gets smarter. Google reports DeepSWE rising to 73.7% (within a point of Claude Opus 5) and OSWorld-2.0 to 59.0%; independently, Artificial Analysis scores it 59 on its Intelligence Index at high reasoning, up from 56 on 3.7 Flash and on par with GPT-5.6 Sol and Grok 4.6. And it does this at a promotional $0.75 / $3.75 per million tokens — roughly 1/15th the sticker of OpenAI/Anthropic flagships — while beating those same flagships on several specialist agent benchmarks (finance, legal, HLE-Verified). The asterisks are real: the figures are vendor-reported, the model reasons harder so per-task costs rise ~30–45%, and it still trails flagships on long-horizon autonomous work (Terminal-Bench 4.0 at 19.1% vs Opus 5's 51.8%). Oh, and the promo price doubles on January 1, 2027. For the next four months, this is the default "workhorse" of the agentic tier.

$Gemini 3.8 Flash review 2026 — Google's agentic workhorse beats flagships on specialist evals at a fraction of the price, AA 59, $0.75/$3.75$

Gemini 3.8 Flash — the Flash-tier model that finally got smarter, not just cheaper

What Is Gemini 3.8 Flash?

Released September 2, 2026 by Google DeepMind, Gemini 3.8 Flash is the latest model in Google's budget "Flash" tier — the third update in about six weeks (3.6 Flash on July 21, 3.7 Flash on August 13). Stable model ID: gemini-3.8-flash. It is available on the Gemini API, Google AI Studio, Gemini Enterprise, and to Google AI Pro/Ultra subscribers, alongside a security-specialized sibling, Gemini 3.8 Flash Cyber.

Where the earlier Flash releases were framed as "fast + cheap" iterations, Google is positioning 3.8 Flash around software engineering, agentic tasks, and multi-step reasoning — and reviewers agree the gains are genuinely about capability this time, not just latency. It launched the same week as a wave of frontier releases (OpenAI GPT-6 Astra, Anthropic Claude Fable 5.1, Meta Muse Spark 1.3), which makes it the value play of a very crowded week. For context on the top of the market, see our Claude Fable 5.1 review and Grok 4.6 review.

What Changed vs Gemini 3.7 Flash

Agentic gains, not cosmetic (Google-reported): DeepSWE v1.1 65.3% → 73.7% (near Claude Opus 5's ~74%); Terminal-Bench 2.1 85.8% → 89.4%; OSWorld-2.0 (computer use) 50.6% → 59.0%; HLE-Verified 53.6% → 54.9%.
Specialist strength: Vals Finance Agent v2 59.0% → 61.4%; Harvey Legal Agent Benchmark 8.8% → 10.0% — enough that it reports beating larger frontier models on finance, legal, HLE-Verified and video understanding at a fraction of the price.
Independent confirmation of the intelligence jump: Artificial Analysis Intelligence Index 59 at high reasoning, up from 56 (3.7 Flash) and 50 (3.6 Flash) — roughly on par with GPT-5.6 Sol and Grok 4.6 (61 with their own effort tiers), below Claude Opus 5's 63. Analysts attribute the jump mainly to agentic behavior and tool use, not raw knowledge.
The trade-off: the model reasons harder, averaging ~48,000 output tokens per task (vs ~3.7 Flash), so per-task costs rise ~30–45% despite identical token prices — ~$0.58 per composite task at high effort vs ~$0.40 for 3.7 Flash.
Verbosity: independent reviewers note it produces roughly 30% more output tokens than 3.7 Flash on the same prompts.

Specifications at a Glance

Item	Value
Vendor	Google DeepMind
Release date	September 2, 2026 (3.6 Flash Jul 21 · 3.7 Flash Aug 13)
Positioning	Agentic "workhorse": software engineering, tool use, multi-step reasoning, cost-sensitive agents
Model ID	`gemini-3.8-flash`
Context window	1,048,576 (1M) input tokens · 65,536 (~64K) output tokens
Modalities	Text, image, video (up to 3h low-res / 1h high-res), audio (up to 9.5h), PDFs → text output
Reasoning	Configurable `low / medium / high` thinking effort; thinking enabled by default
Output speed	~300–350 tokens/s (top-3 of ~200 models on Artificial Analysis); TTFT ~13s due to mandatory reasoning
Tools / features	Function calling, code execution, file search, Search & Maps grounding, structured outputs; computer use in preview
Not included	No image/audio generation, no Live API
Open weights	No (closed; API and Google products)
Availability	Gemini API, Google AI Studio, Gemini Enterprise, Google AI Pro/Ultra

Benchmarks: First Flash That Actually Got Smarter

All Google figures below are launch-day, vendor-reported and have not yet been independently replicated. "High" refers to the thinking effort=high tier.

Benchmark	3.7 Flash High	3.8 Flash High	Frontier context
AA Intelligence Index (independent)	56	59	Ties GPT-5.6 Sol & Grok 4.6 (~61 with their effort tiers); below Claude Opus 5 (63)
DeepSWE v1.1	65.3%	73.7%	Within ~1 pt of Claude Opus 5 (~74%)
Terminal-Bench 2.1	85.8%	89.4%	Strong for the tier
OSWorld-2.0 (computer use)	50.6%	59.0%	Below flagship Opus 5 (75.4%)
HLE-Verified	53.6%	54.9%	Beats several larger flagships (Google-reported)
Vals Finance Agent v2	59.0%	61.4%	Reported above larger frontier models
Harvey Legal Agent (Vals)	8.8%	10.0%	Low in absolute terms (brutally hard eval), but a jump
Terminal-Bench 4.0 (long-horizon)	—	19.1%	Weak spot: vs Claude Opus 5 (51.8%)

Cost-per-task in practice: because output pricing includes thinking tokens and the model reasons harder, Artificial Analysis estimated roughly $0.58 average cost per composite agentic task at high effort (~48K output tokens), vs ~$0.40 for 3.7 Flash — the token price is unchanged, but the model spends more.

Pricing: Cheap for Six Months, Then It Doubles

Tier	Input / 1M	Output / 1M
Promo (through Dec 31, 2026)	$0.75	$3.75
Standard (from Jan 1, 2027)	$1.50	$7.50
Batch	$0.375	$1.875

The promo price is the story — and so is the expiry. At $0.75/$3.75, Gemini 3.8 Flash undercuts essentially every comparable agentic model. For sticker comparison (list prices at time of writing, promo windows vary):

Model	Input / 1M	Output / 1M
Gemini 3.8 Flash (promo)	$0.75	$3.75
Gemini 3.7 Flash	$0.75	$3.75
Grok 4.6 (standard)	$2.00	$6.00
GPT-5.6 Sol	$4.00	$20.00
Claude Opus 5	$5.00	$25.00
Claude Fable 5.1	$10.00	$50.00

Three gotchas that matter for real agent workloads:

It doubles on Jan 1, 2027. Any cost model you build on the $0.75/$3.75 promo should bake in the $1.50/$7.50 price for January. Batch pricing (half) is the lever for high-volume, non-latency-sensitive jobs.
Thinking tokens are billed at output rates. Because reasoning is on by default and the model is verbose (~30% more output tokens than 3.7 Flash), sticker price understates per-task cost by 30–45%.
At roughly 1/15th the per-token sticker of OpenAI/Anthropic flagships, the headline math is compelling — but only if the task is one the model can actually finish (see the weak spots below).

Gemini 3.8 Flash Cyber: The Gated Security Twin

Launched the same day, Gemini 3.8 Flash Cyber shares the same 1M-context architecture but carries "more permissive cyber mitigations," enabling autonomous vulnerability discovery and patching that the standard model refuses. Because of its dual-use nature it is not publicly available: access is gated through Google's Fairwind Program, giving 650+ vetted organizations (governments, critical-infrastructure operators, software maintainers, named security vendors) prioritized access for authorized defensive work only, with multi-factor authentication and no-malware commitments. It is paired with Google's CodeMender tooling to generate, validate, and deploy patches in minutes.

Key reported results (Google and partner figures, not independently audited):

Chrome patching: Google's Chrome Security team reports 3.8 Flash Cyber produced 2.6× more correct patches on real Chrome vulnerabilities than larger, best-in-class commercial models.
Vulnerability discovery: 86.2% on CyberGym (autonomous discovery), beating its predecessor and larger frontier models; over 70% success on Google's internal 20-language benchmark.
Automated patching: CWE-Bench pass@1 of 47.2% vs 47.8% for a leading frontier model — at significantly lower cost.
Speed: Google's Cloud Vulnerability Research team surfaced a critical foundational vulnerability in under two hours — work that normally takes months.
Third-party: Wiz measured +7.5–9.7% higher recall on its internal pentest benchmark at 2.3–5.2× lower cost; Google also cites a significant leap in prompt-injection robustness (Gray Swan).

Fairwind mirrors the same "public model + gated cyber twin" pattern as Anthropic's Project Glasswing (Mythos 5.1) and OpenAI's Project Daybreak. For most readers the takeaway is simple: the standard Gemini 3.8 Flash can already do code-review-style vulnerability hunting — just without Cyber's expanded permissions.

The Honest Take

The real headline is a capability jump at a budget price. For a model family that historically shipped "same smarts, cheaper and faster," 3.8 Flash is the first Flash-tier release where the intelligence number itself moved (AA 56 → 59). On Google's numbers it now sits within a point of Claude Opus 5 on DeepSWE and beats bigger, pricier models on finance, legal, HLE-Verified and video understanding — at roughly 1/15th the per-token sticker. For coding agents, tool-use pipelines and cost-sensitive workloads, that is a genuinely hard combination to beat during the promo window.

The asterisks are the usual ones plus a real behavioral trade-off. Every Google figure here is vendor-run on Google's own harnesses (differing frame-sampling and thinking-level methodology; Google also initially misreported Claude Opus 5's DeepSWE score due to rounding, and some 3.8 figures shipped only as images, not text). Independently, the model still trails flagships on long-horizon autonomous work — Terminal-Bench 4.0 at 19.1% vs Opus 5's 51.8%, OSWorld-2.0 at 59.0% vs 75.4% — and its reasoning-is-on-by-default design means every task burns ~30–45% more tokens than 3.7 Flash. Cheap per token does not automatically mean cheap per task.

Bottom line: if you need the absolute ceiling on long-running autonomous agents, Gemini 3.8 Flash is not it — Opus 5 and the frontier flagships still own that. If you want a fast, genuinely smarter workhorse for coding, tool-calling and specialist agent evals at a fraction of the price — and you ship before the promo doubles on January 1, 2027 — this is currently the strongest value argument in the entire agentic tier.

Pros & Cons

Pros

✅ First Flash-tier release where intelligence meaningfully advanced: AA Index 59 (vs 56 on 3.7 Flash), DeepSWE 73.7% (Google-reported)
✅ Devastating price during the promo: $0.75/$3.75 — roughly 1/15th the per-token sticker of OpenAI/Anthropic flagships
✅ Beats larger frontier models on specialist agent evals: finance (Vals 61.4%), legal, HLE-Verified, video understanding (Google-reported)
✅ Massive 1M context, multimodal input (video/audio/PDF), ~300–350 tok/s output (top-3 on AA)
✅ Full tool belt: function calling, code execution, file search, Search & Maps grounding, structured outputs
✅ The gated Cyber twin shows genuinely state-of-the-art defensive capability (2.6× Chrome patches)

Cons

❌ Vendor-reported benchmarks with no independent replication yet; methodology caveats (frame sampling, image-only figures, an initial Opus 5 DeepSWE misreport)
❌ Weak on long-horizon autonomous work: Terminal-Bench 4.0 19.1% and OSWorld-2.0 59.0% vs Opus 5's 51.8% / 75.4%
❌ Verbose and reason-hard-by-default: ~30% more output tokens, per-task costs ~30–45% above 3.7 Flash
❌ Promo price doubles to $1.50/$7.50 on January 1, 2027
❌ No image/audio generation or Live API; closed weights
❌ Flash Cyber is gated to 650+ vetted defenders — most users cannot access it at all

Who Should Use It (And Who Shouldn't)

Great fit

Coding-agent and tool-use workloads where token cost across many turns dominates
Finance, legal, research and video-understanding pipelines that need frontier-adjacent specialist performance cheaply
Google AI Pro/Ultra or Gemini Enterprise users who get it via their existing subscription
Teams that can ship before Jan 1, 2027 to lock in the promo price (or use batch)

Poor fit

Long-horizon autonomous agents that need the Terminal-Bench / OSWorld ceiling (flagships win there)
Token-frugality purists — lower reasoning settings or Gemini 3.7 Flash may cost less per task
Buyers who only move on independently replicated benchmark numbers
Cybersecurity teams expecting to use the Cyber variant — access is gated via the Fairwind Program

Frequently Asked Questions

When was Gemini 3.8 Flash released?

September 2, 2026, by Google DeepMind — the third Flash-tier update in ~6 weeks (3.6 Flash Jul 21, 3.7 Flash Aug 13). Available on the Gemini API, Google AI Studio, Gemini Enterprise, and Google AI Pro/Ultra.

How much does it cost?

$0.75 input / $3.75 output per million tokens through Dec 31, 2026; $1.50 / $7.50 from Jan 1, 2027. Batch is half. Output pricing includes thinking tokens, so per-task cost runs ~30–45% above the sticker math.

Is it better than Gemini 3.7 Flash?

Yes on intelligence: AA Index 59 vs 56, DeepSWE 73.7% vs 65.3%, OSWorld-2.0 59.0% vs 50.6% (Google-reported). The cost: it reasons harder and emits ~30% more output tokens, raising per-task cost.

What is the context window?

1M input tokens and ~64K output. Multimodal input (text, image, video up to 3h low-res / 1h high-res, audio up to 9.5h, PDFs) → text output. ~300–350 tokens/s, with ~13s time-to-first-token from mandatory reasoning.

How does it compare to Claude Opus 5 / GPT-5.6 Sol?

AA Index 59 ties GPT-5.6 Sol and Grok 4.6, behind Opus 5 (63). Google reports it beats larger flagships on finance, legal, HLE-Verified and video understanding at ~1/15th the sticker; it trails them on Terminal-Bench 4.0 and OSWorld-2.0 long-horizon work. Vendor-reported.

What is Gemini 3.8 Flash Cyber?

A cybersecurity-specialized variant for autonomous vulnerability discovery and patching (2.6× more correct Chrome patches per Google). Not public — gated via the Fairwind Program to vetted defenders. See the Cyber section above.

Is the promo price permanent?

No — $0.75/$3.75 runs only through December 31, 2026, then doubles to $1.50/$7.50. Plan cost models accordingly.

Sources

Disclosure: benchmark figures are launch-day, vendor-reported by Google DeepMind (or partner-measured for Cyber) and not yet independently replicated; pricing and availability are subject to change — confirm on Google's official Gemini documentation.

搜索此博客

ring