Gemini 3.5 Pro

Spec

Attribute	Value
Developer	Google DeepMind
Released	not yet (third deadline missed 2026-07-17; no confirmed date)
Announced	Google I/O, May 19, 2026
Context window	2,000,000 tokens (2M — largest in any production frontier model)
Pricing	~$1.25/$10 per million tokens expected at GA (standard tier); Deep Think TBD — estimates conflict, see Conflicting Reports
License	unknown
Availability	Preview only — Vertex AI enterprise preview + gradual developer API rollout (as of 2026-07); AI Studio whitelist closed 2026-07-12
Status (as of 2026-07-21)	Pending — Gemini 3.6 Flash stopgap shipped (July 21); 3.5 Pro confirmed "on the way" by Google
Expected GA	TBD — prediction markets July 31 (81%) or Aug 7 (73%); stopgap (3.6 Flash) now GA
Architecture	New pre-training cycle (2.5 Pro base discarded; rebuild targets math, SVG, image quality)
Reasoning mode	"Deep Think" (gated to $250/month Ultra tier)
Modality	Text + Images (confirmed); additional multimodal TBD

Release Date

Announced at Google I/O 2026 (May 19, 2026) with a June 2026 general-availability target. As of July 17, 2026: Gemini 3.5 Pro has missed its third consecutive deadline. Google is considering a "Gemini 3.6 Flash" stopgap release while Pro development continues. (Bloomberg) (9to5Google) (source)

Timeline

• May 19, 2026: announced at Google I/O, June GA promised
• June 25, 2026: delay to July announced
• July 7, 2026: expanded developer preview began; delay causes confirmed by testers (token overuse, coding gaps, long-horizon reasoning)
• July 8, 2026: July 17 target date confirmed; full architectural rebuild of 2.5 Pro base disclosed
• July 12, 2026: AI Studio whitelist closed without a public release; GA appeared to have slipped to July 22-28
• July 14, 2026: July 17 GA re-confirmed — multiple sources confirm July 17 as definitive launch target; rebuild complete (source)
• July 17, 2026: ⚠️ THIRD DEADLINE MISSED — July 17 GA did not materialize. Bloomberg and 9to5Google (July 16) confirmed coding performance fell short of GPT-5.6 benchmarks; hallucination rate above Google's internal bar. Stopgap: Google registered "Gemini 3.6 Flash" and "Gemini 3.5 Flash Light" model names — interim releases under consideration. Prediction markets: 81% for July 31 next; 73% for Aug 7. No confirmed date. (Bloomberg) (9to5Google) (TechTimes) (source)

Benchmarks

No public benchmark data released yet (developer preview). Full benchmarks are expected at GA (date TBD after the July 17 miss).

Reported delay reasons (from early enterprise testers + architectural rebuild disclosure, as of July 8, 2026):

• Full rebuild motivation: 2.5 Pro base architecture scrapped — failures were fundamental enough to require a new pre-training cycle
• Excessive token consumption in multi-step reasoning chains (model "thinks too much" on simple tasks)
• Coding performance below the bar Google set at I/O for the Pro tier
• Long-task, multi-step reasoning falling short of I/O-promised benchmarks
• SVG scene generation and image quality below competitors (GPT-5.6, Fable 5)

(source) (source)

Talent Context

The architectural restart coincides with four senior Gemini/DeepMind researchers departing in June 21–27, 2026:

• Noam Shazeer (VP Engineering, Gemini co-lead, Transformer co-author) → OpenAI (June 18)
• John Jumper (Nobel Chemistry 2024, AlphaFold) → Anthropic (June 19)
• Jonas Adler (AI coding effort, pretraining) → Anthropic
• Alexander Pritzel (pretraining) → Anthropic

The Adler/Pritzel departures were the "two more" reported by TheNextWeb in the week following Shazeer and Jumper. (source)

Key Differentiators

• 2M token context window — doubles Gemini 2.5 Pro's 1M context; the largest context window in any production frontier model as of June 2026.
• Deep Think reasoning mode — extended test-time compute, gated to Google One AI Premium Ultra ($250/month) subscribers.
• Multimodal — confirmed text + image support at minimum; expected audio/video consistent with Gemini family architecture.

Positioning

Follows Gemini 3.5 Flash (GA May 19, 2026), which is the speed-optimized variant. Gemini 3.5 Pro is the capability-maximizing variant in the 3.5 family.

Competes directly with:

• Claude Fable 5 — (currently suspended, 2026-06-12)
• GPT-5.5 (standard)
• Claude Opus 4.8 — current Anthropic production model

Sources

• TechTimes (June 6 pre-launch analysis): https://www.techtimes.com/articles/317919/20260606/google-gemini-35-pro-nears-june-launch-2-million-token-context-deep-think-reasoning.htm
• Announcement at Google I/O: https://blog.google/products/gemini/gemini-3-5-pro-io-2026/

Use Cases

TBD at GA. Expected: long-context document analysis, complex reasoning tasks (Deep Think), multi-step agentic workflows.

Compared To

Model	Context	Status
Gemini 3.5 Pro	2M	Developer preview (as of July 2026)
Gemini 3.5 Flash	1M	GA (May 19, 2026)
Claude Fable 5	1M (SWE-bench Pro 80.3%)	Suspended (June 12, 2026)
Claude Opus 4.8	1M	GA
GPT-5.5	TBD	GA

Related

• Google DeepMind — developer
• Gemini 3.5 Flash — sibling model (GA)
• Reasoning Models — Deep Think reasoning mode
• Test-Time Compute (Inference-Time Compute Scaling) — underlying mechanism for Deep Think

Conflicting Reports

• Expected GA pricing: body follows the more recent July reporting — ~$1.25/$10 per million tokens (standard tier), per the July 7 (source) and July 8 (source) snapshots. Conflicting earlier media estimate: ~$15/$60 per million tokens with consumer access via a $20/month Gemini Pro plan (GrowWing Assistant, June 19, 2026). No official pricing has been published.
• Modality: body follows the July 7 snapshot (text + image confirmed; audio/video TBD — source). Conflicting earlier media claim: "full multimodal capabilities confirmed" across text, image, video (GrowWing Assistant, June 19, 2026). No official spec sheet has been published.
• Second missed deadline: TechTimes/9to5Google (July 16) give the missed-deadline chain as June → July 1 → July 17 (source), while July 7–8 reporting described the interim target as a "late July" window that July 17 then replaced (source). Body keeps the three-deadline count without asserting the July 1 date.