Google Pulled Claude From Antigravity’s Free Tier. The Gemini Flash It Kept Beats Opus 4.6 78.65% to 53.50% at Vibe Coding.

What changed in Antigravity on November 2

Antigravity is Google’s agent-first coding tool. Its pricing page lists a $0 plan for individuals with “Unlimited Tab completions” and “Basic weekly rate limits” on agent models.

Until this week, those free agent models included 3 from outside Google: Claude Sonnet 4.6 (thinking), Claude Opus 4.6 (thinking) and OpenAI’s open-weight GPT-OSS-120b. Google’s model docs flag all 3 with the same footnote: “Will be removed on November 2, 2026.”

That leaves Gemini 3.8 Flash as the only reasoning model on the Free and Google AI Plus plans. It’s also the only model the docs list on every plan.

Source: Google Antigravity Docs, Models, opened November 3, 2026. Gemini 3.7 Flash, 3.6 Flash and 3.1 Pro are listed for Enterprise only.

The newer Claude models never reached the free tier. Google lists Claude Sonnet 5.5 and Opus 5.5 as “Available on Google AI Pro for non-trial subscriptions only,” so a Pro trial doesn’t get them either.

The surprise: the model they kept is better at vibe coding

Losing Claude sounds like a downgrade. On the benchmarks Vals.ai has run on both, it mostly isn’t. Gemini 3.8 Flash came out on September 2, 2026, 7 months after Opus 4.6 (February 5) and Sonnet 4.6 (February 17), and it shows.

Sources: Vals.ai leaderboards, Vibe Code Bench v1.1 (updated October 29, 2026), SWE-bench Verified and LiveCodeBench (updated September 1, 2026). Vibe Code Bench runs Gemini 3.8 Flash and Opus 4.6 in OpenHands; Sonnet 4.6 scores 51.48% in OpenHands and 55.77% in Claude Code. Opus 4.6 without thinking scores 57.57%.

The Vibe Code Bench gap is the big one: 25.15 points over Opus 4.6 (Thinking) and 22.88 over Sonnet 4.6 in Claude Code. That benchmark is the closest match to what most people do in a free agent: describe an app and let the model build it.

On SWE-bench Verified the lead shrinks to 1.8 points over Opus 4.6, and Flash costs more per test, $2.19 to $1.22. Close enough that free users fixing bugs in an existing repo won’t feel a big change either way.

The catch: Terminal-Bench 4.0

Terminal-Bench 4.0 is where Gemini 3.8 Flash falls apart. It gives an agent long tasks in a real terminal: set up services, debug builds, chain commands for an hour or more. Vals.ai measures 19.70% for Gemini 3.8 Flash, 34th of 59 models.

Source: Vals.ai Terminal-Bench 4.0 leaderboard, updated October 29, 2026, Mini-SWE-agent harness. Cost per solved test is cost per test divided by accuracy, our arithmetic. Vals.ai hasn’t run Claude Opus 4.6 or Sonnet 4.6 on Terminal-Bench 4.0.

Two numbers stand out. First, both paid Claude 5.5 models score more than 3x higher. Second, Gemini 3.8 Flash is the most expensive model per solved task in that table: $43.91, more than twice Opus 5.5’s $20.26. DeepSeek V4.1 Flash gets the same 19.70% for $0.50 a test, 17x less.

The reason is runtime. Vals.ai lists Gemini 3.8 Flash at 1h47m per test, the longest of any model in the table above. A cheap per-token model that keeps working on a task it can’t finish still runs up a bill.

On Antigravity’s free plan, you don’t see that bill. You see it as quota: Google’s plans page says the rate limits “correlate with the amount of work done by the agent.” Long, failing terminal runs will eat a free user’s weekly allowance fastest.

Gemini 3.8 Flash’s full coding report card

Vals.ai has run Gemini 3.8 Flash on more coding tests than the 4 above. The pattern holds: strong on short, well-defined coding problems, weak on long terminal and operations work.

Source: Vals.ai, Gemini 3.8 Flash model page, opened November 3, 2026. Vals.ai runs it at high reasoning effort.

The 2 Vibe Code Bench rows are worth reading together. On version 1.1, Flash builds a working app 78.65% of the time. On Vals.ai’s separate Vibe Code Bench 1-100, it scores 18.77%. A strong score on one app-building test doesn’t carry over to every other one.

Same story for operations work: SRE Bench comes in at 11.07%, 13th of 35. ProgramBench, at 1.00%, ranks 17th of 64, so low scores are the norm there, not a Flash problem.

What the paid upgrade buys

If the free tier isn’t enough, Antigravity’s next step is Google AI Pro, which adds Gemini 4 Argon and, on non-trial subscriptions, Claude Sonnet 5.5 and Opus 5.5. On Vibe Code Bench v1.1, Argon scores 91.91% for $8.11 a test: 13.26 points over Flash for $1.24 more per test.

Claude Sonnet 5.5 leads that board at 92.39%, but costs $31.25 a test, and Opus 5.5 costs $57.92 for 90.29%. For app building, Argon is the value pick. For the terminal, the Claude 5.5 models lead Terminal-Bench 4.0 and cost less per solved test than Argon.

What free users actually get

Google’s plans page doesn’t publish numbers for any tier. Free and other non-Pro plans get “a meaningful quota, refreshed weekly.” Pro gets “a high, generous quota, refreshed every five hours until the weekly limit is reached.” Ultra gets the highest quota plus “Access to third-party models.”

Pro and Ultra can also keep going past the quota with AI credits billed at Gemini Enterprise rates. The plans page says Antigravity doesn’t support bring-your-own-key, so a free user can’t plug in an Anthropic key to get Claude back.

If you want Gemini 3.8 Flash outside Antigravity, Google’s Gemini API pricing page lists a free tier for it, then $0.75 per million input tokens and $3.75 per million output through December 31, 2026. From January 1, 2027, those become $1.50 and $7.50: double.

What this means if you code with AI

For vibe coding, the free tier didn’t get worse. If you use Antigravity’s free plan to build apps from a prompt, the remaining model beats both Claude 4.6 models on Vals.ai’s app-building test by more than 22 points.

For terminal-heavy agent work, it did. 19.70% on Terminal-Bench 4.0 means about 1 in 5 long terminal tasks finishes. If your work is DevOps, build fixes or multi-step shell sessions, a free Flash model isn’t the tool, in Antigravity or anywhere.

Check the benchmark that matches your work. On Vals.ai, Gemini 3.8 Flash ranks 3rd of 143 on LiveCodeBench and 34th of 59 on Terminal-Bench 4.0. A single “best coding model” ranking hides that.

Free tiers change without a blog post. This change lives in a footnote on a docs page. If a free AI coding tool is part of your workflow, keep a fallback that doesn’t depend on one vendor’s free model list.

Full disclosure: we sell an AI coding agent, and we did not run these benchmarks. MegaBrain’s free plan needs no card and includes the free models, plus your own API keys and local models (Ollama, LM Studio) with no limits from us. See vibe coding with MegaBrain or the plans. MegaBrain is not affiliated with Google, Anthropic, OpenAI or Vals.ai.

Sources

All pages opened November 3, 2026.

• Google Antigravity Docs — Models (plan availability; Claude Sonnet 4.6, Claude Opus 4.6 and GPT-OSS-120b “will be removed on November 2, 2026”)
• Google Antigravity Docs — Plans (quotas, refresh periods, AI credit overages)
• Google Antigravity — Pricing ($0 for individuals, Business from $30 per seat)
• Google AI for Developers — Gemini API pricing (Gemini 3.8 Flash, prices through and after December 31, 2026)
• Vals.ai — Vibe Code Bench v1.1 leaderboard (updated October 29, 2026)
• Vals.ai — Terminal-Bench 4.0 leaderboard (updated October 29, 2026)
• Vals.ai — SWE-bench Verified leaderboard (updated September 1, 2026)
• Vals.ai — LiveCodeBench leaderboard (updated September 1, 2026)
• Vals.ai — Gemini 3.8 Flash model page (all benchmark results)