Leaked Benchmarks Claim Google's Unreleased Gemini 4 Pro Tops Frontier Models

NokiaPowerUser has published what it says are leaked benchmark scores for an unannounced Google DeepMind model called Gemini 4 Pro, claiming it outperforms current and forthcoming flagship models from Anthropic and OpenAI across coding, agentic, and reasoning tasks. The leak adds a Pro variant to Google's Gemini 4 lineup alongside Gemini 4 Argon (mid-October), Gemini 4 Carbon (October 28, 2026), and a Gemini 4 Flash Preview variant referenced in a separate SDK leak.

What the leak claims

The NokiaPowerUser report presents a table comparing Gemini 4 Pro against Claude Opus 5.5 and GPT-6 Astra. The provided sources do not confirm the release status of these comparison models. The claimed scores show Gemini 4 Pro leading on every listed benchmark:

| Benchmark | Gemini 4 Pro (leak) | Claude Opus 5.5 | GPT-6 Astra | |-----------|---------------------|-----------------|-------------| | DeepSWE v1.1 | 88.7 | 74.2 | 74.1 | | Terminal-Bench 2.1 | 95.3 | ~88 | ~88 | | OSWorld 2.0 | 86.8 | 81.8 (partial) | 72.6 (offline partial) | | GDPval-AA (Elo) | 2,064 (v2) | 1,846 (v2.1) | 1,542 (v2.1) | | Context window | 2M tokens | 1M tokens | 1.1M tokens | | Price / 1M tokens (in/out) | $2.25 / $11.25 | $4 / $20 | $10 / $50 |

The leak also states an API pricing structure that undercuts both named rivals, and a 2-million-token context window—roughly double the capacity of the comparison models.

Sources and provenance

The benchmarks appear only in the NokiaPowerUser article, which says the data is "circulating across developer forums and leaderboard tracking communities" and references an anonymous checkpoint labeled gemini-3.8-flash spotted during early stealth testing. No primary source (Google, Anthropic, OpenAI, or a benchmarking organization) is cited. The article does not provide screenshots, repository links, or reproducible evaluation logs.

A separate NokiaPowerUser report described a Google SDK leak referencing gemini-4-flash-preview mapped to a new Gemma 4 tokenizer family, suggesting Google is actively integrating fourth-generation architecture into public tooling. That leak did not mention a "Pro" variant.

What Google has actually released

Google DeepMind announced Gemini 4 Carbon on October 28, 2026, calling it "our strongest model with state-of-the-art cybersecurity and long horizon capabilities." Carbon scored 59 on the Artificial Analysis Intelligence Index v5.0, 88.2% on SEC-bench Pro (vulnerability detection), and 54.8% on ExploitGym (exploitation). Access is limited to a subset of cyber defenders via the Fairwind Program, with a gradual rollout planned. Pricing at launch matches the earlier Gemini 4 Argon: $2 / $10 per million input/output tokens (introductory), rising to $4 / $20 after the promotional period.

Google's blog post announces Carbon and references Argon, which it says "was launched two weeks ago." A separate Safety and Security report for Argon was previously published. As of the time of writing, Google has not publicly announced a Gemini 4 Pro model; its blog post and SDK leaks reference only Argon, Carbon, and a Flash Preview variant.

The comparison models are also unreleased

The leak compares Gemini 4 Pro against Claude Opus 5.5 and GPT-6 Astra. The provided sources do not cover Anthropic's or OpenAI's announcement histories, so their release status is not confirmed by the available reporting.

Benchmark interpretation caveats

• Scores from uncontrolled, unreleased checkpoints can reflect "benchmaxxing"—tuning for specific test sets rather than general capability.
• The leak notes OSWorld scores for the rivals as "partial" and "offline partial," suggesting different evaluation conditions.
• Pricing for unreleased models is speculative; Google's own introductory pricing for Carbon/Argon is lower than the leak's claimed Opus 5.5 and GPT-6 Astra prices, but standard Carbon pricing ($4/$20) matches the leak's Opus 5.5 figure.
• No independent replication, third-party audit, or open evaluation harness is referenced.

Industry context

The leak arrives as developers anticipate OpenAI's DevDay and as Google continues a rapid cadence: Argon (mid-October), Carbon (October 28), SDK references to Gemini 4 Flash (early November), and now a claimed Pro variant. NokiaPowerUser also recently published a rumor about a Gemini 4 Xenon successor to Carbon, citing unnamed "insider leaks" without primary evidence.

Third-party rankings show a mixed picture for Carbon. It tops the Artificial Analysis Intelligence Index but ranks #2 on the Coding Agent Index behind Claude Opus 5.5, and #4 on WebDev Arena behind Opus 5.5, GPT-6.1 Astra, and Sonnet 5.5 [5].

What remains unknown

• Whether a model called Gemini 4 Pro exists in any form beyond internal checkpoints.
• Whether the benchmark numbers are genuine, current, or representative of a shipping product.
• Google's intended release timeline, pricing, or access policy for any Pro variant.
• How the claimed performance translates to real-world developer workflows outside the cited benchmarks.

Bottom line

A single outlet has published unattributed benchmark figures for three unreleased models, one of which (Gemini 4 Pro) has not been acknowledged by its maker. The numbers are striking but unverified. Until Google, Anthropic, or OpenAI confirm model existence and release evaluation data under reproducible conditions, the leak should be treated as a rumor—not a result.