Between July and early September 2026, nine labs shipped real models: and the internet invented several more. This is a dated, verified map of what actually exists as of 11 September 2026, what status each release is in, and what each one is genuinely useful for if your job is marketing. It also includes a section most roundups skip: the models that do not exist but rank well anyway. If you've searched this topic recently, you've probably read about at least one of them.
I write this as a practitioner, not an analyst. I care about which of these I can put into a production workflow next Monday.
Key Takeaways
- Three frontier releases landed in a nine-day window: Claude Fable 5.1 (1 Sep), Gemini 3.8 Flash (2-3 Sep) and GPT-6 Astra (3 Sep).
- GPT-6 Astra and Claude Fable 5.1 have identical headline pricing at $10/M input and $50/M output. The differentiator is caching, not list price.
- Google's strategy is Flash-first: four Flash models in 106 days, and Gemini 3.5 Pro has been announced but has not shipped.
- Open-weight options got genuinely good: Qwen3.8 27B is small enough to self-host, and Z.AI's GLM-5.3 Flash shipped open-weight in August.
- Mistral repositioned away from general chat toward verticals, and raised €3B at a €21B+ valuation in September 2026.
- Grok 5 does not exist. Every result claiming otherwise is speculative SEO content.
- Google's Project Astra is a research prototype and has nothing to do with OpenAI's GPT-6 Astra. The name collision causes real confusion.
The Complete Release Table
Everything below is verified against primary sources as of 11 September 2026. Where a date is contested across sources, I say so rather than picking one.
| Lab | Model | Date | Status | Pricing (in/out per M) | Marketer relevance |
|---|---|---|---|---|---|
| OpenAI | GPT-6 Astra | 3 Sep 2026 | Released | $10 / $50 | Flagship. Strategy, analysis, agent orchestration. Fast mode: 2x speed, 2x price. Off by default for enterprise admins initially. |
| Anthropic | Claude Fable 5.1 | 1 Sep 2026 | Released | $10 / $50, cache reads $0.25 | Flagship. Best economics for repeated-context workflows. Anthropic claims ~25% cost reduction typical, up to 45% agentic. |
| Anthropic | Claude Mythos 5.1 | 1 Sep 2026 | Restricted | n/a | Same underlying model as Fable, different safeguards. Vetted cybersecurity and life-sciences professionals only. Marketers cannot access it. |
| Anthropic | Opus / Sonnet / Haiku | Ongoing lineup | Available | Varies | Haiku-class is the volume workhorse: product descriptions, ad variants, localisation. |
| Gemini 3.8 Flash | 2-3 Sep 2026 | Released | Not restated here | Google's best available model. Roughly 10th on the Artificial Analysis index. Cheap, fast, very high volume. | |
| Gemini 3.5 Pro | Announced | Not shipped | n/a | "Coming soon" as of September 2026. Do not plan around it. | |
| Meta | Muse Spark 1.3 + "Contributor" | 2 Sep 2026 | Released | n/a | Powers Meta Muse, the consumer agent launched 8 Sep. Relevance is distribution, not API access. |
| Meta | Muse Spark 1.1 | Jul 2026 | Released | n/a | Predecessor. |
| xAI | Grok 4.6 | 12 Aug 2026 | Released | n/a | On a 1.5T-parameter foundation. Strong on real-time social context. |
| xAI | Grok 4.5 | 8 Jul 2026 | Released | n/a | Predecessor on the same foundation. |
| xAI | Grok 4.7 | Announced | Not released | n/a | Announced only, as of September 2026. |
| Qwen | Qwen3.8 Flash | 26 Aug 2026 | Released, open-weight | Self-host | High-volume generation without per-token API cost. |
| Qwen | Qwen3.8 27B | 2 Sep 2026 | Released, open-weight | Self-host | 27B is small enough to self-host on sane hardware. The most practically deployable open model in this window. |
| Z.AI | GLM-5.3 Flash | 26 Aug 2026 | Released, open-weight | Self-host | Another credible open-weight option for volume work. |
| Mistral | Agentic Search | Aug 2026 | Released | n/a | Vertical product, not a general chat model. |
| Mistral | Robostral Navigate, Leanstral 1.5 | Jul 2026 | Released | n/a | Robotics and efficiency verticals. |
| Mistral | Mistral OCR 4 | Jun 2026 | Released | n/a | Document extraction. Genuinely useful for content ops. |
| DeepSeek | V4 Pro / V4 Flash | Dates conflict | Released | n/a | Both exist. Primary sources disagree on timing (April vs August), so I won't state a date. |
Reading the table
Two things jump out. First, the frontier cluster: 1, 2-3 and 3 September. Three labs shipped flagship-tier releases inside three days, which is either coincidence or a very expensive game of chicken.
Second, the open-weight column is no longer a consolation prize. Qwen3.8 27B at a self-hostable size changes the cost structure of any workflow that generates thousands of short outputs.
What Is NOT Real
This is the section I'd most like you to read. The volume of confidently-written content about models that don't exist is, as of September 2026, genuinely a problem for anyone trying to plan.
Grok 5 does not exist
There is no Grok 5. xAI shipped Grok 4.5 on 8 July 2026 and Grok 4.6 on 12 August 2026, both on a 1.5-trillion-parameter foundation. Grok 4.7 has been announced but not released.
Every article, benchmark table and "Grok 5 vs GPT-6" comparison you find is speculative content written to capture anticipatory search volume. Some of it invents benchmark scores. If you're briefing a team or a client, this is the single most likely place for a false fact to enter your deck.
Gemini 3.5 Pro has not shipped
Google announced Gemini 3.5 Pro. It is still "coming soon" as of September 2026. Google's best available model is Gemini 3.8 Flash, which sits around 10th on the Artificial Analysis index.
The numbering is confusing, 3.8 Flash is newer than the unshipped 3.5 Pro, and a lot of coverage gets this backwards. If someone tells you they're building on 3.5 Pro today, they aren't.
Google Project Astra is not GPT-6 Astra
Two entirely different things share a name:
- Google Project Astra is a research prototype. It is not a shipping product.
- OpenAI's Astra is GPT-6, released 3 September 2026, a shipping flagship model.
I've seen this conflation in published articles, in vendor decks and in at least one investor memo. When you see "Astra", check whose.
No genuinely new frontier labs entered this window
Between July and September 2026, no new lab entered the frontier tier. Any content claiming a surprise new competitor should be treated with heavy scepticism until you can find a primary source. Well-funded does not mean frontier.
OpenAI: GPT-6 Astra
Released 3 September 2026 at $10/M input and $50/M output. There's a Fast mode at roughly double the speed for double the price.
The enterprise default
Worth knowing if you work inside a larger organisation: GPT-6 Astra was off by default for enterprise admins initially. If your team "doesn't have access", that's likely an admin toggle rather than a rollout queue.
Benchmark claims
OpenAI has published performance claims for Astra. Those are vendor self-reported figures, and I'd attribute them as such in any document you circulate. Independent evaluation takes weeks; as of 11 September 2026 the picture is still mostly the lab's own.
Anthropic: Claude Fable 5.1 and the Restricted Mythos
Released 1 September 2026. Same headline pricing as Astra: $10/M in, $50/M out.
The caching change is the real news
Cache reads dropped roughly 75% to $0.25/M. Anthropic claims this yields about a 25% cost reduction on typical workloads and up to 45% on agentic tasks.
For marketing specifically, this matters more than it sounds. Most serious marketing AI work involves feeding the same large context, brand guidelines, tone of voice, product catalogue, past campaigns, into every single call. That's exactly the pattern cache reads optimise. If your workflow re-sends 30,000 tokens of brand context a thousand times a month, this pricing change is a line item you'll notice.
Mythos 5.1 is not for you
Claude Mythos 5.1 is the same underlying model as Fable with different safeguards, restricted to vetted cybersecurity and life-sciences professionals. Marketers cannot access it. I mention it only because it appears in Anthropic's lineup listings, alongside Fable, Opus, Sonnet and Haiku, and people keep asking how to get it.
Google: The Flash-First Strategy
Google shipped four Flash models in 106 days. That cadence is a strategy, not an accident.
The price trajectory
Gemini 3.7 Flash launched at half the per-million-token price of 3.6 Flash: three weeks after it. Read that again. Three weeks, half the price.
If you're building cost models on Google's pricing, build them to be re-run monthly. The direction is aggressively downward.
What Flash-first means for retrieval
A Flash-first strategy implies Google is optimising for volume and latency at consumer scale. Given that the Gemini app reached 1 billion monthly users, the fastest-growing product in Google's history, that makes sense. It also shapes how Gemini retrieves and answers, which has real consequences for brand visibility inside it.
xAI, Qwen, Z.AI, Mistral and DeepSeek
xAI
Grok 4.5 (8 July) and 4.6 (12 August) on a 1.5T-parameter foundation. Grok 4.7 announced, not shipped. The genuine differentiator remains real-time access to social conversation, which is narrowly but genuinely useful for trend monitoring.
Qwen and Z.AI
Qwen3.8 Flash (26 August) and Qwen3.8 27B (2 September), both open-weight. Z.AI's GLM-5.3 Flash (26 August), also open-weight.
The 27B is the one I'd actually evaluate. Self-hostable size means the marginal cost of a product description goes from "a fraction of a cent" to "electricity", and at genuine volume that's a different business case.
Mistral
Mistral deliberately stepped away from competing on general chat. Instead: Agentic Search (August), Robostral Navigate and Leanstral 1.5 (July), Mistral OCR 4 (June). They raised €3B in a Series D at a €21B+ valuation in September 2026.
There is no credible Mistral general flagship in this window, and that appears to be on purpose. The vertical products are worth a look, OCR 4 in particular if you process a lot of documents.
DeepSeek
V4 Pro and V4 Flash both exist. Primary sources conflict on release timing, with April and August both cited. I'm not going to pick one. If a date matters to your decision, verify it directly before publishing it.
How I'd Use This Map
For production volume
Flash and Haiku-class models, or a self-hosted Qwen3.8 27B. Product descriptions, ad variants, localisations, meta descriptions at scale.
For strategy and analysis
GPT-6 Astra or Claude Fable 5.1. Positioning work, competitive analysis, campaign architecture, editorial QA.
For agentic workflows
Fable 5.1's cache economics tilt the maths here, especially for anything that re-reads the same brand context on every step.
For document processing
Mistral OCR 4 is purpose-built and does one thing well.
The Meta-Lesson About This Category
Four frontier-adjacent releases in nine days, a 50% price cut in three weeks, announced models that haven't shipped and invented models that rank. This is what an immature market looks like. The practical implication for a marketing team is to build workflows that are model-agnostic at the interface layer. Don't hard-code a model name into fifty prompts. Route by task type, swap the underlying model when the economics change, and re-check your assumptions every quarter.
I'll update this map as things ship. Everything here is dated 11 September 2026 for a reason.
FAQ
What are the latest AI models as of September 2026?
The three most recent frontier releases are Claude Fable 5.1 (1 September), Gemini 3.8 Flash (2-3 September) and GPT-6 Astra (3 September). Meta's Muse Spark 1.3 also shipped 2 September, and Qwen3.8 27B on 2 September as an open-weight model.
Is Grok 5 out yet?
No. Grok 5 does not exist. xAI's latest released model is Grok 4.6 from 12 August 2026, with Grok 4.7 announced but unreleased. Content claiming to review or benchmark Grok 5 is speculative.
Has Gemini 3.5 Pro been released?
No. It has been announced and is described as coming soon. Google's best available model as of September 2026 is Gemini 3.8 Flash.
What's the difference between Google Astra and OpenAI Astra?
They're unrelated. Google's Project Astra is a research prototype, not a product. OpenAI's Astra is GPT-6, a released flagship model from 3 September 2026.
Which is cheaper, GPT-6 Astra or Claude Fable 5.1?
Headline pricing is identical: $10/M input, $50/M output. Fable 5.1 cut cache reads roughly 75% to $0.25/M, so for workflows that repeatedly send the same context, Fable is cheaper in practice. Anthropic claims ~25% typical savings and up to 45% on agentic work.
Can I use Claude Mythos 5.1?
Not if you're a marketer. It's restricted to vetted cybersecurity and life-sciences professionals. It shares the underlying model with Fable but carries different safeguards.
What's the best open-weight model right now?
Qwen3.8 27B, released 2 September 2026, is the most practically useful because 27B parameters is small enough to self-host on reasonable hardware. Qwen3.8 Flash and Z.AI's GLM-5.3 Flash, both from 26 August, are also open-weight.
Why did Mistral stop competing on general chat?
Mistral repositioned toward verticals, agentic search, robotics, OCR, rather than a general flagship. They raised €3B at a €21B+ valuation in September 2026, so the market appears to accept the strategy.
When did DeepSeek V4 release?
Primary sources conflict, citing both April and August 2026. Both V4 Pro and V4 Flash exist; I'd verify the date directly before relying on it.
How often should I re-check this landscape?
Quarterly at minimum, monthly if AI cost is a material line item. Gemini 3.7 Flash launched at half the price of 3.6 Flash three weeks after it. Assumptions expire fast.
Sources and Further Reading
- OpenAI, for GPT-6 Astra specifications and pricing.
- Anthropic: for the Claude 5.1 lineup, pricing and caching details.
- Google's blog, for Gemini release notes and the billion-user milestone.
- TechCrunch, for launch coverage across labs.
If you want help turning this landscape into a workflow that actually ships work, get in touch. I've spent 4+ years in marketing helping edtech and startup brands grow organically, including the Masai School run from 26K to 117K on Instagram and 50K to 160K on LinkedIn. See the work and reach me through the contact form at younusfardeen.com.