Removing Gemini 3.1 Pro from Antigravity (agy) for paying Google AI Pro subscribers—while leaving us stranded on Gemini 3.8 Flash—shows a staggering disconnect between your product management and actual software engineering workflows.
Whatever contaminated synthetic benchmarks you used to justify replacing a dense Pro model with a distilled Flash variant, they do not survive contact with real-world codebases. We all understand that 3.8 Flash cuts your inference costs by 70%, but passing off an aggressive cost-optimization downgrade as an “upgrade” is pure marketing delusion.
In production CLI usage, Gemini 3.8 Flash is fundamentally broken:
- Blatant ingestion hallucinations: You cannot even trust it to read a local file or technical document. It routinely claims to have read a document you explicitly pointed it to, only to hallucinate its contents and fabricate APIs, constraints, or sections out of thin air. An agentic CLI that lies about reading context is worse than useless—it is actively destructive.
- Chronic context amnesia: It ignores half the explicit constraints in the initial prompt, loses track of system architecture within two turns, and casually overwrites working, tested code with broken regressions.
- The
--effort highplacebo: Telling senior engineers to just force--effort highis laughable. Making a heavily compressed Flash model burn more thinking tokens does not magically give it the attention capacity or world-knowledge density of a Pro model; it just takes longer to output the exact same shallow, junior-level slop.
The product positioning here makes zero sense:
- The “Pro” branding absurdity: How do you justify charging for a tier literally named Google AI Pro while restricting your flagship developer CLI exclusively to a Flash model?
- The Web vs. CLI incoherence: As an AI Pro subscriber, I already have access to Gemini 4 Argon in the Gemini web interface. The compute entitlement is already part of my plan. Artificially blocking Argon in
agyjust to force CLI users into an overpriced Ultra upsell—while forcing us to copy-paste code back and forth from a browser tab like it’s 2023—is hostile UX design.
Like many engineers here, Claude Code is already my primary daily driver. The only reason I maintained a paid Google AI Pro subscription was to have a dense, frontier-class, non-Anthropic model directly in my terminal for architectural second opinions and cross-verification. I do not need agy to run Claude (Claude Code does it natively and far better), and asking 3.8 Flash to audit Claude Opus is like asking a rushed intern to review a Staff Engineer’s design doc.
If Google AI Pro subscribers do not get access to Gemini 4 Argon (or a restored dense Pro model) inside agy within the next 7 days, I am canceling my subscription and reallocating that budget straight to OpenAI’s Codex CLI (GPT-5.5 / gpt-5.3-codex). Stop treating paying Pro developers like free-tier beta testers.