Gemini 3.8 Flash as a Cursor and Cline BYOK Backend in 2026: Same $0.75/M Price as 3.7, Better Scores, and 108 Days Until It Doubles

By AICoderScope Team · 13 min read
geminigooglecursorclineapicostbyokpricing

TL;DR: Gemini 3.8 Flash shipped September 2, 2026 at the exact price of 3.7 Flash — $0.75 input / $3.75 output per million tokens — with better coding scores everywhere Google published one (73.7% vs 65.3% on DeepSWE v1.1). The upgrade is free; the deadline isn’t. Both models ride the same introductory-pricing clock, and on January 1, 2027 the rate doubles to $1.50/$7.50.

Gemini 3.8 FlashClaude Sonnet 5DeepSeek V4.1 Flash
Best forVolume agent loops, the current price-performance defaultHardest multi-file changes, max fidelityAbsolute lowest cost, US-hours batch work
Input / Output per 1M$0.75 / $3.75 through Dec 31 → $1.50 / $7.50$2.00 / $10.00 (permanent)$0.15 / $0.60 off-peak; $0.30 / $1.20 peak
DeepSWE v1.173.7% (Google-run)53.8% (Google-run comparison, from the 3.7 launch set)not published
The catchPrice doubles Jan 1, 2027; 13.3s time-to-first-token; Cursor BYOK currently erroringTokenizer counts ~30% more tokens per taskRepriced twice in four weeks; peak/off-peak billing
Context window1M in / 65,536 out1M128K

Honest take: If 3.7 Flash is your Cursor or Cline backend, change the model string to gemini-3.8-flash today — same bill, measurably better agent runs. Then put January 1 in your actual calendar: the day this model’s price doubles is the day DeepSeek V4.1 Flash at $0.15/$0.60 becomes 12× cheaper, and that’s a decision worth making with a diary reminder instead of a surprise invoice.

What did Google ship on September 2?

Gemini 3.8 Flash is Google’s fourth Flash model in under four months — most recently 3.6 Flash on July 21, 3.7 Flash on August 13, and now 3.8 Flash on September 2, 2026. Artificial Analysis, which tracks every release independently, scores it 59 on its Intelligence Index and puts it on the intelligence-vs-cost Pareto frontier — the point where nothing published is both smarter and cheaper.

The launch pattern broke in one important way: there was no price cut this time. When 3.7 Flash launched, it halved 3.6 Flash’s price. 3.8 Flash launches at the identical $0.75/$3.75 introductory rate, on the identical December 31, 2026 expiry — Google upgraded the model and left the clock running. The spec sheet also holds steady: 1,048,576-token context window, 65,536-token output ceiling, and the stable model ID gemini-3.8-flash on the Gemini API, AI Studio, and Vertex AI.

One sibling worth knowing about even though you can’t use it: Gemini 3.8 Flash Cyber, a security-tuned variant that Google says produced 2.6× more correct vulnerability patches in Chrome than much larger commercial models. It’s gated behind the invite-only Fairwind Program for governments, critical-infrastructure operators, and open-source maintainers — not on the public API, so nothing to configure in your coding stack.

Is Gemini 3.8 Flash actually better than 3.7 Flash for coding?

On every number Google published, yes — and the DeepSWE jump is the headline. Google’s launch benchmarks, all run by Google:

BenchmarkGemini 3.8 FlashGemini 3.7 Flash
DeepSWE v1.1 (long-horizon SWE)73.7%65.3%
Terminal-Bench 2.1 (terminal agents)89.4%not the headline metric at 3.7 launch
SWE-Bench Pro61.6%60.4%

The 73.7% on DeepSWE v1.1 deserves the same asterisk we attached to 3.7 Flash’s 65.3% three weeks ago: DeepSWE is Datacurve’s independent benchmark, but this run was executed and reported by Google, not posted on the official leaderboard. What makes it notable is the neighborhood — Claude Opus 5’s published leaderboard score sits at roughly 74%, so Google is claiming its $0.75 speed-tier model now scores within a point of the frontier model that anchors the escalation tier in our Opus 5 review. Treat that as a claim to test on your own repo, not a settled ranking. The SWE-Bench Pro move (+1.2 points) is the more honest signal: real but incremental.

The independent numbers cut both ways. Artificial Analysis clocks output at 305 tokens per second — third fastest of the 195 models it tracks — but measures a 13.30-second median time-to-first-token with reasoning on, against a 2.99-second median for reasoning models in its price tier. In an agent loop that fires dozens of sequential calls, throughput dominates and 3.8 Flash flies. In interactive chat where you’re waiting on the first character, that 13 seconds is a real, feelable lag. Match the model to the workload accordingly.

What does Gemini 3.8 Flash cost — and when does the price double?

Verified against Google’s published pricing, September 14, 2026:

TierInput / 1MOutput / 1MNotes
Standard (through Dec 31, 2026)$0.75$3.75Introductory rate
Standard (from Jan 1, 2027)$1.50$7.50Published at launch — not a rumor
Batch / Flex$0.375$1.875Half of standard, non-interactive
Priority$1.35$6.75Latency-guaranteed tier
Cache read$0.075—Through Dec 31; agent loops with stable system prompts benefit most

There’s also a genuine free tier: 3.8 Flash is usable in AI Studio and through the Gemini API’s free tier with rate limits, which is enough to validate your Cline config before attaching a card. The standing trade applies — Google can use free-tier content for product improvement, so keep client code on a paid key.

The number that matters for planning is 108: as of today, that’s how many days of introductory pricing remain. Because 3.7 and 3.8 Flash share the same expiry, switching between them buys you zero extra runway. Budget forecasts built on $0.75/$3.75 are wrong from January 1, and Google published that fact on launch day.

How do you set up Gemini 3.8 Flash in Cursor?

The short version: use Cursor’s native integration today, because BYOK is currently broken for this specific model.

Native is trivial. Cursor added Gemini 3.8 Flash shortly after launch and publishes a dedicated model page for it — open Settings → Models, enable Gemini 3.8 Flash in the picker, and it routes through Cursor’s own infrastructure against your plan’s usage.

BYOK is where this launch stumbled. Multiple reports on the Cursor forum (early September 2026) describe the same failure: with a Google AI Studio key configured, selecting Gemini 3.8 Flash throws “Unauthorized User API key” — while 3.7 Flash on the same key works without issue. The key isn’t the problem; the Cursor-side model routing is. Until Cursor ships a fix, you have two working configurations:

  1. Native 3.8 Flash — plan usage, no key management, works now.
  2. BYOK 3.7 Flash — direct Google billing at the same per-token price, one benchmark generation behind.

Before blaming either tool, a 30-second sanity check proves your key and the model string both resolve:

curl -s "https://generativelanguage.googleapis.com/v1beta/models/gemini-3.8-flash:generateContent?key=$GEMINI_API_KEY" \
  -H 'Content-Type: application/json' \
  -d '{"contents":[{"parts":[{"text":"Reply with exactly one word: ready"}]}]}'

A working key returns JSON whose candidates[0].content.parts[0].text is ready and whose modelVersion echoes gemini-3.8-flash. If that succeeds and Cursor still errors, it’s the Cursor bug — stop debugging your key.

The standing caveat is unchanged: Cursor Tab autocomplete always runs on Cursor’s own models. BYOK affects Chat and Agent only.

How do you set up Gemini 3.8 Flash in Cline?

Cline’s first-party Google Gemini provider takes an AI Studio key directly, and the free tier means you can test before paying:

  1. Open Cline settings, set the API provider to Google Gemini.
  2. Paste your AI Studio key.
  3. Pick gemini-3.8-flash from the model dropdown. If your build’s dropdown hasn’t caught up — model catalogs across the tool ecosystem are still adding 3.8 entries as of mid-September — update the extension first, and only then hand-type the ID.

One config landmine is new with this release: the MINIMAL thinking level is gone. Gemini 3.8 Flash accepts LOW, MEDIUM (the default), and HIGH — and explicitly setting thinking_level to MINIMAL returns an API validation error rather than silently falling back. If you carried a minimal-thinking config forward from an older Flash model, your requests will start failing on the new model string. Google’s guidance from 3.7 carries over: leave it on MEDIUM for agent sessions, drop to LOW only when interactive latency matters more than plan quality — which, given the 13-second reasoning TTFT, it sometimes will.

What does an agent workload actually cost — today and in January?

On our standard session shape — 20K input tokens, 7K output, the same math used across our backend reviews:

BackendPer sessionPer 1,000 sessions/mo
DeepSeek V4.1 Flash (off-peak / US hours)$0.007$7
DeepSeek V4.1 Flash (peak)$0.014$14
Gemini 3.8 Flash (intro, through Dec 31)$0.041$41
Gemini 3.8 Flash (standard, from Jan 1)$0.083$83
Claude Sonnet 5 ($2/$10, ~30% tokenizer overhead)$0.143$143

The DeepSeek rows changed since our last comparison, and in DeepSeek’s favor. On September 10, 2026 — eight days after 3.8 Flash launched — DeepSeek replaced V4 Flash outright with V4.1 Flash under the deepseek-flash model ID: a 552B-parameter MoE with MIT-licensed open weights, priced at $0.15/$0.60 off-peak and $0.30/$1.20 during peak windows (weekdays 01:00–04:00 and 06:00–10:00 UTC). That undercuts even the pre-hike V4 Flash rates, and US working hours still fall entirely off-peak. It’s also the second DeepSeek repricing in four weeks, continuing a pattern of short-notice changes that makes it a floor-price option you monitor rather than a rate you budget around.

Read the table twice — once for now, once for January. Today, Gemini 3.8 Flash costs about 6× DeepSeek’s US-hours rate and buys you first-party serving, a 1M context window, and the strongest published agent scores in its price class. From January 1, the same gap becomes 12×, and the “just use Flash for everything” default stops being obvious. If your volume is high enough that $83 vs $7 per thousand sessions matters, the January decision is between DeepSeek’s floor pricing and going local — Nemotron 3.5 Lightning runs free in Cline, our sister site’s VRAM-based local model guide covers what your GPU can serve, and open-weight options live at aifoss.dev. The full market picture is in our cost comparison across every tier.

When should you NOT switch to Gemini 3.8 Flash?

Four cases where the right answer is something else:

  • You’re on Cursor BYOK and can’t switch to native routing. The “Unauthorized User API key” bug means 3.8 Flash doesn’t work over BYOK right now. Stay on 3.7 Flash — same price, same expiry, no functional regression — and retest after the next Cursor update.
  • Your workload is interactive chat, not agent loops. A 13.3-second reasoning time-to-first-token is the wrong shape for conversational use. Sonnet 5 or a lower thinking level serves that better, and at LOW you’re trading away the benchmark gains you switched for.
  • Your bill must survive January unchanged. Any budget built on $0.75/$3.75 has a 108-day fuse. If you can’t absorb a 2× on this line item, start migrating volume work to DeepSeek V4.1 Flash or a local model now, while it’s a choice rather than an emergency.
  • You need output beyond 65K tokens in one shot. The output ceiling is unchanged from 3.7 Flash, and very large single-shot generations still truncate. Incremental agent writes rarely hit it; whole-file rewrites of big files do.

And one non-case: waiting for 3.9. On current cadence Google ships a new Flash roughly every four weeks, and each has been a drop-in string change. Deferring a free upgrade because another free upgrade might come is how you end up two generations behind in December with the price doubling anyway.

Verdict

Gemini 3.8 Flash is the rare launch where the decision takes one sentence: same price, same clock, better model — move the string. It replaces 3.7 Flash as the default volume backend for Cline today and for Cursor via native routing until the BYOK fix lands. The two real dates on your calendar are whenever Cursor patches BYOK, and January 1, 2027, when this model stops being a $41-per-thousand-sessions default and becomes an $83 line item you should re-decide deliberately — against a DeepSeek that just got cheaper and a local-model floor that costs electricity. We’ll re-run this math when the clock hits.

FAQ

What’s the exact model ID? gemini-3.8-flash on the Gemini API, AI Studio, and Vertex AI. Stable GA identifier, no suffix. Migrating from 3.7 Flash is a one-string change.

Did the price change from Gemini 3.7 Flash? No. Both cost $0.75 input / $3.75 output per million tokens, and both rates expire on the same day — December 31, 2026 — before doubling to $1.50/$7.50. Switching models does not extend the discount window.

Why does Gemini 3.8 Flash fail with my API key in Cursor? A known Cursor-side bug, reported on the Cursor forum in early September 2026: BYOK requests to 3.8 Flash return “Unauthorized User API key” while the same key works on 3.7 Flash. Use Cursor’s native Gemini 3.8 Flash integration, or stay on 3.7 Flash over BYOK until it’s fixed.

Is the 73.7% DeepSWE score independently verified? Not yet. DeepSWE v1.1 is Datacurve’s independent benchmark, but the 73.7% run was executed and reported by Google at launch. It would put 3.8 Flash within about a point of Claude Opus 5’s published leaderboard score — credible directionally, provisional until Datacurve posts it.

Can I set thinking to minimal for faster agent loops? No — MINIMAL is not a valid thinking level on 3.8 Flash and returns an API validation error. The supported levels are LOW, MEDIUM (default, recommended for coding and agents), and HIGH.

What is Gemini 3.8 Flash Cyber and can I use it? A security-tuned variant for vulnerability discovery and patching, available only through Google’s invite-only Fairwind Program for governments, critical-infrastructure operators, and software maintainers. It is not on the public API.

Sources

Last updated September 14, 2026. Pricing and features change frequently — Gemini Flash introductory pricing ends December 31, 2026; verify current rates before committing a budget.

Was this article helpful?

Know which coding tool is worth paying for

Hands-on comparisons of AI coding assistants and what each one costs to run — including the local-model path. Sent only when something changes. Unsubscribe anytime.