Gemini 3.7 Flash as your Cursor and Cline backend in 2026: 65.3% DeepSWE, $0.75/$3.75 per million tokens, and a price that expires December 31

By AICoderScope Team · 13 min read
geminigooglecursorclineapicostbyoksetup-guide

TL;DR: Gemini 3.7 Flash went GA on August 13, 2026 — three weeks after 3.6 Flash — at an introductory $0.75 input / $3.75 output per million tokens, exactly half its predecessor. Google reports 65.3% on DeepSWE v1.1 (up from 49%) and 43.6% on Cognition’s FrontierCode 1.1, ahead of Claude Sonnet 5’s 42.7%. The catch: the price doubles back to $1.50/$7.50 on January 1, 2027.

Gemini 3.7 FlashClaude Sonnet 5DeepSeek V4-Flash (post-hike)
Best forVolume agent loops, the new cheap defaultHardest multi-file changes, max fidelityAbsolute lowest cost, off-peak batch work
Input / Output per 1M$0.75 / $3.75 through Dec 31 → $1.50 / $7.50$2.00 / $10.00 (now permanent)$0.44 / $1.32 peak; $0.22 / $0.66 off-peak
DeepSWE v1.165.3% (Google-run)53.8% (Google-run comparison)not listed
The catchIntro price expires Dec 31, 2026Tokenizer counts ~30% more tokens per task371% output price hike Aug 16; peak/off-peak billing
Context window1M in / 65,536 out1M128K

Honest take: If 3.6 Flash is your volume backend, switch to 3.7 Flash today — it scores higher on every published benchmark and costs half as much until year-end, with zero config changes beyond the model string. Keep Sonnet 5 for the changes that must land right the first time, and re-run this math on January 1 when the discount dies.

Three weeks, 16 points, half the price

Google released Gemini 3.7 Flash on August 13, 2026 — twenty-three days after Gemini 3.6 Flash went GA. That cadence is the story as much as the model is. The Flash tier is now iterating faster than most teams re-evaluate their tooling, and Google is pricing each new release to force the question: 3.7 Flash launched at $0.75 per million input tokens and $3.75 per million output, half of 3.6 Flash’s $1.50/$7.50, with context caching at $0.075 per million.

Read the fine print before you celebrate. This is introductory pricing, and it has a published expiry: on January 1, 2027, the rates double to $1.50/$7.50 — exactly what 3.6 Flash costs today. Caching rises to $0.15 per million on the same date. So the honest framing is not “Google cut Flash prices in half.” It’s “Google is giving you a 4.5-month discount to migrate, after which 3.7 Flash costs what 3.6 Flash always did.” Since the model is better at the same eventual price, migrating is still the right call — but budget forecasts built on $0.75/$3.75 will be wrong in January, and Google has already told you so.

The spec sheet holds steady from 3.6 Flash: 1,048,576-token context window, 65,536-token output ceiling, knowledge cutoff of March 2026, multimodal input (text, image, video, audio, PDF) with text output, and function calling, search grounding, and computer use as built-in tools. The model ID is gemini-3.7-flash — a stable GA identifier, no -exp or -preview suffix. It’s live in the Gemini API, AI Studio (including the rate-limited free tier), Android Studio, Google Antigravity, and the Gemini Enterprise platform.

The benchmark read: a real jump, with the usual asterisk

Google’s launch numbers, all run by Google:

BenchmarkGemini 3.7 FlashGemini 3.6 FlashReference points
DeepSWE v1.1 (long-horizon SWE)65.3%49.0%GPT-5.6 Terra 69.6%, Sonnet 5 53.8% (Google-run)
FrontierCode 1.1 Main (code quality)43.6%34.4%Sonnet 5 42.7%, GPT-5.6 Terra 41.3%
AutomationBench30.4%17.0%—
WebDev Arena1588 Elo1538 Elo—

Two of these deserve unpacking. DeepSWE v1.1 is Datacurve’s independent benchmark — 113 hand-authored, long-horizon tasks across 91 active open-source repositories in five languages, graded by committing the agent’s code into a clean, isolated environment. When we covered 3.6 Flash’s 49% three weeks ago, that was already the best score-per-dollar on the board. A 16.3-point jump in one iteration is the kind of gain that usually takes a major version. The asterisk is the same one as last time: the 65.3% comes from Google’s own evaluation run on Datacurve’s harness, not from the official leaderboard, where Claude Opus 5 leads the published snapshot at 73.6% pass@1. Google’s own comparison set puts GPT-5.6 Terra ahead at 69.6% and Sonnet 5 well behind at 53.8% — but every one of those cross-model numbers is one vendor grading its competitors, so treat the ranking as provisional until Datacurve posts it.

FrontierCode 1.1 is newer and worth knowing about: it’s Cognition’s benchmark (the Devin company), and it grades code quality — not just “does it pass,” but whether the output is code you’d merge. Scoring 43.6% there, nominally ahead of Sonnet 5’s 42.7%, is the most aggressive claim in the launch material, because Sonnet’s whole value proposition is exactly that kind of production fidelity. A 0.9-point edge on a vendor-run eval is not proof Flash writes better code than Sonnet. It is evidence the gap has narrowed enough that you should run both on your own repo before paying 2.7× more per output token.

What Google says changed under the hood, for what it’s worth: more careful multi-step planning, more reliable tool calls, and better recovery when an agent run hits a roadblock — the failure modes that actually burn money in Cursor and Cline loops, because a flubbed tool call costs you the whole retry.

Wiring it into Cursor

Cursor moved faster on this release than any previous Gemini launch. There was no multi-day picker lag like 3.6 Flash in July: Cursor already publishes a dedicated model page for Gemini 3.7 Flash confirming native support, with access to the full agent toolset — file search, directory reads, web search, vision-enabled file reads, edit application, terminal execution, and browser control.

So the setup for most people is: open Settings → Models, find Gemini 3.7 Flash in the picker, enable it, done. Native routing counts against your Cursor plan and needs no key management.

The BYOK route still matters if you want direct cost visibility or you’re on a plan where native usage is constrained:

  1. Get a key at Google AI Studio (aistudio.google.com) — free, no card required.
  2. In Cursor: Settings → Models → API Keys, paste it into the Google API Key field.
  3. Add gemini-3.7-flash as the model name if it isn’t listed, then select it for Chat or Agent.

Before pointing either tool at the key, a 30-second sanity check from the terminal confirms the key works and the model resolves:

curl -s "https://generativelanguage.googleapis.com/v1beta/models/gemini-3.7-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 a JSON body whose candidates[0].content.parts[0].text is ready and whose modelVersion field echoes gemini-3.7-flash. A 404 here means the model string is wrong; a 400 with API_KEY_INVALID means the key is — both cheaper to find in curl than mid-agent-run.

The standing caveat applies: Cursor Tab autocomplete always runs on Cursor’s own models. A Google key (or any BYOK backend) changes Chat and Agent only.

Wiring it into Cline

Cline’s first-party Google Gemini provider takes an AI Studio key directly — no OpenAI-compatibility shim:

  1. Open Cline settings, set the API provider to Google Gemini.
  2. Paste your AI Studio key.
  3. Pick gemini-3.7-flash from the model dropdown. If the dropdown hasn’t caught up, update the extension before hand-typing the ID — typing against a stale build is how people end up silently routed to an older model.

One genuine change from the 3.6 Flash guidance: thinking levels. Flash exposes low, medium, and high, and where Google’s previous positioning pushed low/minimal for tool loops, the 3.7 documentation now names medium — the default — as the recommended setting for complex code and agentic use cases, with low reserved for latency-critical work and high for the hardest reasoning problems. In practice: leave it on medium for Cline agent sessions and only drop to low when you’re doing rapid interactive edits where response latency matters more than plan quality. That’s a reversal of the advice we gave for 3.6 Flash, and it comes from Google’s own model docs, not our inference.

The cost math changed twice this month

The competitive landscape this launch drops into looks nothing like it did when our 3.6 Flash guide went up on July 25 — two pricing events rewrote it.

First: Anthropic canceled the Sonnet 5 price increase. The $2/$10 per million rate, originally announced as introductory through August 31 with a scheduled jump to $3/$15, is now the permanent standard price per Anthropic’s pricing docs. Every cost comparison that assumed $3/$15 from September 1 should be read at $2/$10 indefinitely — we’ve corrected the math in the 3.6 Flash article to match.

Second: DeepSeek raised prices on August 16. V4-Flash, the long-standing cheapest credible backend at a flat $0.14/$0.28, moved to peak/off-peak billing: $0.44 input / $1.32 output per million during peak hours (01:00–04:00 and 06:00–10:00 UTC), and $0.22/$0.66 off-peak. That’s a 371% output increase at peak. One useful accident of time zones: US working hours fall entirely in DeepSeek’s off-peak window, so American developers effectively pay $0.22/$0.66 — still the cheapest option going, but the gap to first-party Google pricing collapsed from 10× to roughly 3×.

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

BackendPer sessionPer 1,000 sessions/mo
Gemini 3.7 Flash (intro, through Dec 31)$0.041$41
DeepSeek V4-Flash (off-peak / US hours)$0.009$9
DeepSeek V4-Flash (peak)$0.018$18
Gemini 3.7 Flash (standard, from Jan 1) / 3.6 Flash today$0.083$83
Claude Sonnet 5 ($2/$10, ~30% tokenizer overhead)$0.143$143

Sonnet 5’s effective cost includes the ~30% extra tokens its newer tokenizer counts for the same code. DeepSeek remains the raw-cost floor, but you’re now managing a clock-dependent bill and a model that just demonstrated it will reprice on short notice — and 3.7 Flash’s reported DeepSWE score puts it in a different capability tier. If your monthly agent bill needs to be $0 rather than $41, the local route is the answer: Nemotron 3.5 Lightning runs free in Cline, and our sister site’s VRAM-based local model guide covers what your GPU can actually serve. Open-weight self-hosting options live at aifoss.dev.

Where it breaks

The 65K output ceiling can truncate very large single-shot generations, though incremental agent writes rarely hit it. The hardest cross-cutting refactors are still where the gap to Opus 5 (73.6% on the official DeepSWE board) lives, and a failed agent run costs review time no per-token discount recovers. And the January 1 price doubling is a real planning trap for teams: a $41/month-per-developer line item silently becoming $83 across a 20-person team is a $10K annual surprise if nobody diaried it.

If you’re on GitHub Copilot rather than BYOK: Gemini 3.7 Flash started rolling out in Copilot’s model picker the same day it went GA — August 13 — continuing the same-day pattern 3.6 Flash set in July. Check how model choice interacts with Copilot’s credit billing before assuming it’s covered by your flat rate.

Verdict

Gemini 3.7 Flash replaces 3.6 Flash as the default answer for volume agent work, and the decision requires no agonizing: it is better on every published number and costs half as much until December 31, after which it costs exactly what you’re paying now. Move the model string today. The harder question — whether it replaces Sonnet 5 for work you’d previously escalate — deserves a week of side-by-side runs on your own codebase before you believe a 0.9-point vendor-run FrontierCode edge. Our cost comparison across every tier has the standing market numbers, and the Opus 5 review covers the escalation tier this model still doesn’t reach.

FAQ

What’s the exact model ID? gemini-3.7-flash on the Gemini API, AI Studio, and Vertex AI. Stable GA identifier — no suffix. Migration from 3.6 Flash is literally changing that one string.

When does the introductory pricing end? December 31, 2026. From January 1, 2027, rates double to $1.50 input / $7.50 output per million tokens, and cache pricing rises from $0.075 to $0.15 per million. Google published the expiry at launch.

Is the 65.3% DeepSWE score independently verified? Not yet. DeepSWE v1.1 is an independent Datacurve benchmark with a reproducible, isolated-grading harness, but the 65.3% run was executed and reported by Google. The official published leaderboard is led by Claude Opus 5 at 73.6% pass@1. Directionally credible, provisionally ranked.

Is there a free tier? Yes — AI Studio’s free tier includes Gemini 3.7 Flash with rate limits. The standard trade applies: Google can use free-tier content for product improvement, so keep client code on a paid key.

Should I set thinking to low for agent loops like on 3.6 Flash? No — this changed. Google’s 3.7 Flash docs recommend the default medium thinking level for complex code and agentic use cases. Use low only for latency-critical interactive work, and high for the hardest reasoning tasks.

Does BYOK Gemini change Cursor Tab autocomplete? No. Tab always runs on Cursor’s own models. A Google API key affects Chat and Agent modes only.

Sources

Last verified August 18, 2026. Pricing and model availability change frequently — the introductory rate in this article expires December 31, 2026; check the official pricing page before committing a team budget.

Was this article helpful?