GPT-6 “Sol” soon? Gemini 4.0 Pro checkpoint, DeepSeek V4.1 Flash, Nano Banana 2.5 and more

The week’s real signal: releases, checkpoints and model IDs

The working headline matches the subject: GPT-6 “Sol” soon, Gemini 4.0 Pro checkpoint, DeepSeek V4.1 Flash, Nano Banana 2.5 and more. The current state is not one clean launch event but a cluster of model signals: Google appears to be testing a stronger Gemini Pro checkpoint, DeepSeek has officially shipped V4.1 Flash, OpenAI has fresh product releases but only rumor-level evidence for GPT-6 Sol, and Google’s image pipeline may have a “Nano Banana 2.5” successor moving through arena-style testing [1].

That distinction matters. In today’s AI news cycle, a visible model identifier, a benchmark screenshot, and a community test can all look like a launch. They are not the same. DeepSeek’s V4.1 Flash is confirmed by official release notes and API documentation; the Gemini 4.0 Pro and Nano Banana 2.5 items remain checkpoint sightings and arena speculation; GPT-6 Sol is still best treated as a possible upcoming tier rather than a public OpenAI release [1][2][3].

Google’s Gemini Pro checkpoint: promising, but not yet a product

The Google thread starts with a reported new Gemini Pro checkpoint, described as the first notable Pro-line movement in some time and potentially connected to the future Gemini 4.0 Pro family [1]. The most-circulated demo is a zero-shot SVG peacock generation that reportedly consumed around 24,000 tokens and took roughly six minutes at high reasoning effort, producing an output that testers described as stronger than several competing advanced systems [1].

The Japanese recap of the same WorldofAI segment preserves the same framing: a new Gemini Pro checkpoint has appeared, the model ID is said to be publicly visible, and the previous project checkpoint was reportedly called Argon 160, which had also surfaced in arena testing under the Gemini 3.8 Flash name [5]. That does not prove that Gemini 4.0 Pro is ready, priced, or even named for release. It does suggest that Google is iterating in the background across the Flash and Pro tracks, possibly using public or semi-public evaluation surfaces before a wider rollout [1][5].

The meaningful technical clue is not just that the SVG looked good. It is that a long, structured, visual-code task tests planning, instruction persistence and token-budget discipline. If the report is accurate, the checkpoint may be improving in the area where Google needs credibility most: not raw demo sparkle, but fewer failures across long-form reasoning and generation [1].

Spicy Mayo and the Nano Banana 2.5 question

The image-model subplot is less firm but still relevant. A mystery Google model reportedly appeared in model arenas under the codename Spicy Mayo, with testers speculating that it could be Nano Banana 2.5, a lighter update, or another branch of Google’s image stack [1][5]. In one comparison against GPT-Image 2.5 on a Minecraft camel prompt, OpenAI’s model was judged more accurate to the game interface, while Spicy Mayo was described as more visually appealing [1].

That is exactly the kind of anecdote that should be useful but not overread. Image models often trade off semantic accuracy, aesthetic polish, text rendering, UI fidelity and prompt obedience. A single Minecraft-style comparison can reveal a direction of travel, but it cannot establish a general ranking. The safer takeaway is that Google appears to be testing a follow-up to its Nano Banana image line while OpenAI’s image models are the comparison target in community tests [1][5].

DeepSeek V4.1 Flash is the confirmed launch

DeepSeek is the hard news in this cycle. The company officially introduced DeepSeek-V4.1-Flash on September 10, 2026, calling it the smallest model in a new architecture family and emphasizing native visual understanding, higher throughput and lower cost [2]. The model is available through the DeepSeek API as deepseek-flash, while the older deepseek-v4-flash and deepseek-v4-flash-vision-exp names temporarily route to the new model for compatibility [2][3].

DeepSeek says V4.1 Flash is a 552-billion-parameter mixture-of-experts model using a causal encoder-decoder architecture, with roughly 8 billion active parameters during input processing and 16 billion during output generation [2][4]. That asymmetric design is important because modern agent workloads often read far more than they write: repositories, logs, browser pages, documents and tool outputs. A model that reduces prefill and cache cost can be economically significant even before benchmark bragging begins [2][4].

The other headline is cache compression. DeepSeek says the new model’s KV cache uses one quarter of the HBM and one eighth of the SSD storage of the previous generation, a claim aimed directly at long-context agent economics [2]. Yotta Labs’ analysis frames the same shift bluntly: V4.1 Flash is cheaper and stronger for many API users, but its roughly 510 GB FP8 checkpoint makes self-hosting a node-scale problem rather than a casual two-GPU deployment [4].

Benchmarks: impressive, but workload-specific

DeepSeek’s own changelog lists standout scores including GPQA Diamond 90.9, Codeforces rating 3471, Terminal-Bench 2.1 at 90.6, DeepSWE v1.1 at 74.2, CyberGym at 88.1, HLE with tools at 63.9, and Automation-Bench at 54.8 [3]. The company says testing puts V4.1 Flash ahead of DeepSeek V4 Pro on performance, cost, speed and total runtime, and it plans to route all deepseek-v4-pro traffic to V4.1 Flash from 04:00 UTC on September 14, 2026, until V4.1 Pro is released [2][3].

That is a major product decision. It effectively collapses the old Pro-versus-Flash distinction, at least temporarily. But the nuance is essential: Yotta’s breakdown notes that V4.1 Flash looks especially strong on agentic coding, terminal work and tool-use benchmarks, while knowledge-heavy no-tool tasks such as HLE and GPQA do not show a clean across-the-board victory over V4 Pro [4]. In practical terms, teams using DeepSeek for coding agents may see an upgrade and a price cut; teams using V4 Pro for deep knowledge work should run their own evaluations before the routing switch [4].

GPT-6 “Sol”: likely tier, not confirmed release

The OpenAI portion is the most tempting to overstate. The 8news and Japanese recaps both describe GPT-6 Sol as reportedly spotted in public-facing systems or discussed as part of a tiered rollout after GPT-6 Astra [1][5]. The plausible structure is simple: Astra remains the premium frontier tier, while Sol could become the faster or cheaper six-series model, analogous to how prior OpenAI naming separated high-end and everyday tiers [1].

But the confirmed OpenAI news in the same 72-hour window points elsewhere. OpenAI announced a new Data agent in ChatGPT Work on September 10, letting employees connect company data, ask business questions, and generate answers, dashboards and actions while respecting existing permissions [6]. It also launched GPT-Live-1 in the API, a full-duplex voice model designed to listen and speak simultaneously while delegating deeper reasoning and actions to backend agents and tools [7]. Those are real product moves. GPT-6 Sol, by contrast, remains a leak-and-watch item until OpenAI publishes a model card, API entry or announcement [1][6][7].

Math rumors need the strongest caveat

The most explosive claims in the cycle concern possible AI progress on famous unsolved mathematics problems, including discussion around the Hodge conjecture and Birch and Swinnerton-Dyer conjecture [1][5]. These claims should be treated as unverified. A lab being “near verification” is not the same as a proof being public, peer-reviewed, and accepted by the mathematical community. If AI systems are contributing to Millennium Prize-level work, that will become one of the biggest scientific stories of the decade; for now, it is still rumor, not a settled result [1][5].

The bottom line

The current story is a split screen. DeepSeek has made the concrete move: V4.1 Flash is official, multimodal, aggressively priced, open-weight, and scheduled to absorb V4 Pro API traffic [2][3]. Google appears to be warming up stronger Gemini Pro and image checkpoints, but the evidence is still pre-release [1][5]. OpenAI is shipping enterprise data and voice infrastructure while the GPT-6 Sol rumor mill tries to infer the next model tier from sightings and naming patterns [1][6][7].

For users, the practical advice is straightforward: test DeepSeek V4.1 Flash now if you rely on V4 Pro or coding agents; treat Gemini 4.0 Pro and Nano Banana 2.5 as watchlist items; and do not budget workflows around GPT-6 Sol until it is officially listed. The frontier is moving fast, but not every breadcrumb is a launch.

Sources from the last 72 hours

1. [1]GPT-6 'Sol' Soon! Gemini 4.0 Pro Checkpoint, DeepSeek v4.1 Flash, Nano Banana 2.5, & More! AI NEWS!Sep 11, 2026, 6:15 AM UTC
2. [2]Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient.Sep 10, 2026, 12:00 AM UTC
3. [3]DeepSeek V4.1 Flash: Pricing, Specs, V4 Pro Routing, and How to Access It (2026)Sep 10, 2026, 12:00 AM UTC
4. [4]最新AIニュース：Gemini、DeepSeek、GPT-6 Solの動向Sep 11, 2026, 12:00 AM UTC
5. [5]Now everyone can put data to workSep 10, 2026, 12:00 AM UTC
6. [6]Build more natural voice experiences with GPT‑Live‑1 in the APISep 10, 2026, 12:00 AM UTC
7. [7]Change Log | DeepSeek API DocsSep 10, 2026, 12:00 AM UTC

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.