Four major AI labs launch new models in the first week of September 2026

In the first week of September 2026, OpenAI, Google, Anthropic and Meta released new models in near-simultaneous fashion. The launches followed one another within two to three days and were each accompanied by revised pricing, access arrangements and specialised variants.

The rapid succession underscores how intense competition among the major labs has become. Each model targets its own combination of use cases and cost structures, ranging from heavy mathematical reasoning to agentic workflows with low token consumption.

GPT-6 Astra: OpenAI's new flagship model with restricted cybersecurity access

OpenAI launched GPT-6 Astra on 3 September 2026 as a limited preview for approved organisations. The following day, 4 September, broad availability followed for ChatGPT Plus, Pro, Business and Enterprise users and via the OpenAI API, Microsoft Azure and AWS Bedrock.

According to OpenAI, the model scores 97.6% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 100% on ExploitBench. It is also the first OpenAI model to reach the 'Critical' threshold under the company's own Preparedness Framework, the category reserved for the most severe cybersecurity risks. For that reason, OpenAI has temporarily restricted the most advanced cybersecurity features to the Daybreak Access programme for a selected group of testers.

The standard API price is $10 per million input tokens and $50 per million output tokens. For prompts exceeding 272,000 input tokens, long-context rates of $20 and $75 apply. Regional data-residency endpoints carry a 10% surcharge.

Gemini 3.8 Flash: Google's third Flash release in six weeks

Google introduced Gemini 3.8 Flash on 2 September 2026, the third Flash generation in six weeks. The model scores 90.8% on Terminal-Bench 2.1 and emphasises software engineering, agentic tasks and multi-step reasoning in specialised domains.

Alongside the standard version, Google released Gemini 3.8 Flash Cyber, a variant focused on detecting software vulnerabilities and generating patches. Access to that variant is reserved for 'trusted defenders' through Google's Fairwind Programme.

The introductory price for the standard model is $0.75 per million input tokens and $3.75 per million output tokens, valid through 31 December 2026. From 1 January 2027, those rates double to $1.50 and $7.50 respectively. The model is available via Google AI Studio and the Gemini Enterprise Agent Platform.

Claude Fable 5.1: Anthropic cuts cache costs by 75%

Anthropic opened the series on 1 September 2026 with Claude Fable 5.1. According to Anthropic, the model is intended for long-running and high-stakes tasks that take hours and span multiple applications, such as extensive codebases or complex knowledge-work assignments.

Released alongside it was Claude Mythos 5.1, the same underlying model with lighter safety constraints, accessible to vetted organisations through a trusted-access programme.

The base rates for Fable 5.1 remain unchanged from the previous generation. However, Anthropic reduced the cache-read price by 75%. The company says this translates to approximately 25% lower costs for typical workloads and up to 45% lower for heavy agentic use. Fable 5.1 is available via the Claude API, AWS, Google Cloud and Microsoft Azure.

Muse Spark 1.3: Meta's more efficient agentic model

Meta Superintelligence Labs (MSL), the unit built around Alexandr Wang following Meta's $14.3 billion investment in Scale AI, published Muse Spark 1.3 on 2 September 2026. The model supports text, images, video and files as input and has a context window of over one million tokens.

Compared with Muse Spark 1.2, the model requires approximately 20% fewer tool calls and 25% fewer tokens for the same tasks, according to Meta's internal benchmarks. The standard price is $1.25 per million input tokens and $4.25 per million output tokens, identical to its predecessor. Developers who agree to have their prompts used for model training pay a significantly reduced 'contributor' rate of $0.10 and $0.20 per million tokens.

Muse Spark 1.3 is accessible via Muse Code, Meta's own coding environment, and through the Meta Model API.

Pricing changes at OpenAI and Microsoft Foundry

In addition to the new models, OpenAI adjusted pricing for two variants in the GPT-5.6 family in late July 2026. GPT-5.6 Luna became 80% cheaper and now costs $0.20 per million input tokens and $1.20 per million output tokens. GPT-5.6 Terra fell 20% to $2 and $12. Rates for GPT-5.6 Sol remained unchanged, though that variant received a Fast mode that operates 2.5 times faster at double the price.

As of 1 September 2026, Microsoft Foundry adjusted its pricing for EU Data Zone deployments and regional processing outside the United States. Those locations are now subject to so-called locality premiums; global prices remain unchanged. Microsoft also launched a new APAC Data Zone for customers in Asia-Pacific. The adjustment makes the geographic location of inference an explicit consideration in architecture and cost decisions.

What this means for builders and investors

The concentrated series of launches shows that the major labs are accelerating their release cadence and increasingly shipping specialised variants for specific use cases, including cybersecurity and agentic applications. At the same time, price reductions on existing models and aggressive introductory rates are putting pressure on the cost structures of anyone building on API calls.

For teams building agentic systems, the efficiency improvements in Muse Spark 1.3 and the cache-read reduction in Fable 5.1 are directly relevant to their bills. The introduction of locality premiums at Microsoft Foundry makes cloud architecture a considerably more tangible cost factor for European and Asian customers.