Gemini 3.5 Pro — Rebuilt for Code, Still Unconfirmed

Google DeepMind scrapped the original Gemini 3.5 Pro base model after engineers found it failed at recursive tool-calling and SVG generation, then rebuilt it from updated training data — and according to Bloomberg, those rebuilt results were “disappointing.” The model targeted July 17 for general availability after slipping from the June window Sundar Pichai promised at I/O on May 19. As of today, no official model card, API listing, or pricing confirmation exists. Every spec you’ve read about this model comes from third-party reporting, not Google.

TL;DR

• What: Google rebuilt Gemini 3.5 Pro from scratch after coding failures, targeted July 17 GA — no official launch confirmed
• Why it matters: The rebuild confirms coding quality was structurally broken — not a fine-tuning miss, a pre-training failure
• Context window: Reported 2M tokens (double Flash’s 1M), but unconfirmed by Google
• Action: Do not reroute agent traffic from Flash to Pro until independent benchmarks run on the actual GA build

What Happened

Pichai telegraphed the slip on stage at I/O when he promised “next month” for general availability. That was May 19. June came and went. Bloomberg reported that in late June, Google updated the training data to improve coding skills, but the results were disappointing. The model went through at least one additional training cycle after that failure before reaching its July 17 target — a target that, as of this writing, still lacks official confirmation from Google.

This isn’t a garden-variety delay. A full pre-training restart means the model architecture or data mix was fundamentally inadequate for the coding tasks Google benchmarked internally. The June model and the July model share a name but potentially share very little else. When you scrap pre-training, you’re not patching — you’re admitting the foundation was wrong.

The reported specs are substantial on paper: a 2-million-token context window (double Gemini 3.5 Flash’s confirmed 1M), Deep Think reasoning mode, and estimated pricing around $15 per million input tokens and $60 per million output tokens. But none of these have been confirmed via an official Google source. No model card. No API documentation listing. No pricing page. As of July 13, third-party reporting confirmed that no gemini-3.5-pro listing appeared in the public Gemini API docs.

Every concrete spec for Gemini 3.5 Pro — context window, pricing, Deep Think availability — comes from leaked information and third-party reporting, not an official Google announcement. You are planning around a leak, not a signed launch post.

Why This Matters

The coding failure mode is the real story, and it matters specifically because of how Gemini 3.5 Flash has positioned itself. Flash already ships with confirmed benchmarks: 76.2% on Terminal-Bench 2.1, 83.6% on MCP Atlas, 84.2% on CharXiv Reasoning, running at $1.50 per million input tokens and $9 per million output tokens. Those are real numbers from a real model card. Flash is a known quantity for agentic coding workloads.

Pro was supposed to be the step up — the model you route complex multi-file refactors and deep reasoning tasks to when Flash isn’t enough. But if the rebuild couldn’t satisfactorily close the coding quality gap that triggered the restart in the first place, what exactly is the premium buying you? Estimated pricing puts Pro at roughly 10x Flash per output token. That’s a meaningful cost multiplication, and it only makes sense if Pro delivers materially better results on the tasks that matter to your pipeline.

The 2M context window, if it holds, is a genuine structural differentiator. Nothing else in production doubles Flash’s 1M. Claude Fable 5 and Opus 4.8 cap at around 200K tokens. For codebase-level tasks — full-repo refactors, architectural analysis across dozens of files — a 2M window changes what’s computationally possible. But prior Gemini generations showed degradation in coherence beyond 200–500K tokens in third-party testing. A bigger window that loses the thread at 300K is a marketing spec, not a capability.

Deep Think compounds the evaluation problem. It’s gated to the Ultra subscription tier — $99.99 for the entry tier and up to $200/month for the top tier. Standard Pro subscribers at $19.99/month get the base model without reasoning mode. This two-tier structure means any benchmark comparison against GPT-5.6 Sol or Claude Fable 5 needs to specify which Pro is running. A model comparison that doesn’t account for this pricing gate is comparing different products under the same name.

When evaluating Gemini 3.5 Pro for agent routing, specify whether you’re comparing the base model ($19.99/month Pro tier) or the Deep Think-enabled model (Ultra at $99.99–$200/month). They are functionally different products with different price-performance profiles.

The competitive context shifted while Google delayed. GPT-5.6 Sol launched publicly on July 9. Grok 4.5 opened to the public the same day. Both shipped with published model cards and confirmed pricing before Gemini 3.5 Pro reached GA. Google arrives to a field that already has production-tested answers — and arrives without its own official documentation.

The talent signal adds a structural concern. In the week of June 21–27, four senior Gemini researchers announced departures: Noam Shazeer left for OpenAI, while John Jumper, Jonas Adler, and Alexander Pritzel joined Anthropic. Four senior departures during a rebuild, while the GA target was slipping, doesn’t invalidate the model. But it does raise the platform-stability question for teams making 12-month architecture commitments. The people who built the thing are no longer maintaining it.

The Take

I’ve been watching this since Pichai’s on-stage groan moment at I/O. The rebuild confirms what I suspected: Gemini 3.5 Pro’s original coding capabilities were structurally broken, not just undertrained. A full pre-training restart is a severity level that doesn’t happen because “the docs could be better.” And Bloomberg’s reporting that the rebuilt results were still disappointing means we’re potentially looking at a model that shipped despite not meeting its own internal bar — or one that went through yet another cycle we don’t know about.

Here’s my concrete position: do not reroute agent traffic from Flash to Pro until you’ve seen independent benchmarks on the actual GA build. Not Arena leaks. Not internal previews. The released, versioned, API-accessible model. Flash is a known quantity with published scores and predictable pricing. Pro is a promise backed by a rebuild backed by a delay backed by a team that’s hemorrhaging senior talent.

The 2M context window is the one thing that could justify the switch — if coherence holds across the full window, which no one outside Google has tested. Until someone runs a 1M+ token coding task through the GA model and publishes results, that context window is a spec sheet number, not a capability.

Treat this as watchlist confirmed, routing decision pending. The model may turn out to be excellent. But betting your agent stack on a July 17 promise that missed a June 30 promise first is not engineering — it’s hope. And if you’re making cost comparisons against Claude, make sure you’re comparing the same tier of Pro — the pricing gate between base and Deep Think makes this two different products wearing one name.

Related

• Gemini 3.5 Flash — The Agentic Coding Model: Flash’s confirmed benchmarks and pricing — the baseline Pro needs to beat
• Gemini CLI Cutover — Antigravity Migration Day: The CLI infrastructure Pro will need to integrate with
• Claude Sonnet 5 Tokenizer Cost Trap: How tokenizer differences distort model cost comparisons — relevant for Pro vs. Claude pricing analysis