In October 2026, the top of the AI market is a race between Anthropic, OpenAI and Google, with open-weight models from DeepSeek and Qwen priced far below all of them. For most Australian small and medium businesses, the practical point is simple. The cheaper tier of model is often good enough, so test a few models on your own work and avoid tying your business to one vendor.
Key takeaways
- Anthropic released Opus 5.5, Sonnet 5.5 and Haiku 5.5 between 22 September and 7 October 2026.
- OpenAI’s GPT-6 Astra rolled out in early September. Google’s Gemini 3 Pro, released in November 2025, has had only one update since, and Gemini 4 has been signalled.
- Open-weight models cost a small fraction of frontier prices per token.
- Agents, computer use and the MCP standard are the trends to watch.
- Do not lock into one model. Test on your own tasks before you pay for the top tier.
Anthropic: Opus, Sonnet, Haiku, Fable and Mythos
Anthropic’s newsroom lists three releases in recent weeks. Opus 5.5 came out on 22 September, Sonnet 5.5 on 28 September and Haiku 5.5 on 7 October. Anthropic says Sonnet 5.5 is 30% faster and up to 30% cheaper than Sonnet 5 for most work. It says Opus 5.5 is 40% cheaper to run than Opus 5 and “performs at the level of Claude Fable 5.1 on most work”.
Fable 5.1 and Mythos 5.1 were announced on 1 September 2026. According to Tecnoblog, Fable 5.1 is priced at US$10 per million input tokens and US$50 per million output tokens. Mythos 5.1 is restricted to participants in Project Glasswing.
Glasswing launched on 7 April 2026 around Claude Mythos Preview, which is not generally available. Anthropic lists partners including AWS, Apple, Google, Microsoft, NVIDIA, CrowdStrike, JPMorganChase and the Linux Foundation. On 6 October, Anthropic expanded its Cyber Verification Program, and Glasswing members moved to a “Specialized Access” tier.
OpenAI, Google and Meta
OpenAI’s GPT-6 Astra began rolling out around 3 September 2026 to limited organisations, then to paid ChatGPT tiers. OpenAI lists API pricing of US$10 and US$50 per million tokens in and out. OpenAI reports a score of 72.6% on OSWorld 2.0, a test of computer use. That figure is the company’s own.
For a more neutral comparison, Epoch AI’s Capabilities Index of 16 September 2026 put GPT-6 Astra at 166, Claude Fable 5.1 at 164 and GPT-5.6 Sol at 162. Epoch also reports Fable 5.1 leading on its software engineering sub-index.
Google’s Gemini 3 Pro, released in November 2025, has been updated only once since, according to InfoWorld. Google announced three Gemini Flash models in July 2026. InfoWorld also reports that Google DeepMind’s head, Koray Kavukcuoglu, said Gemini 4 should arrive much earlier than the end of this year.
Meta launched Muse Spark on 8 April 2026. It is the first proprietary frontier model from Meta Superintelligence Labs. Meta has said it hopes to open-source future Muse versions.
Open-weight models and the price gap
DeepSeek announced V4 on 24 April 2026 with open weights. Its API pricing is US$1.74 input and US$3.48 output per million tokens for V4-Pro, and US$0.14 and US$0.28 for V4-Flash. DeepSeek’s published scorecards give its top V4-Pro-Max setting 80.6% on SWE-bench Verified, a coding test.
The Qwen3.5 medium models arrived on 24 February 2026 under the Apache 2.0 licence. They have a native context window of 262,144 tokens and support 201 languages.
The price spread is large. Against Fable 5.1’s list price, DeepSeek V4-Flash is roughly 70 times cheaper on input and about 180 times cheaper on output. That is simple arithmetic on the published figures. It does not mean the two models do the same job equally well. It does mean you should check whether a cheaper model handles your task.
The June export-control episode
Anthropic states that it received a US export-control directive on 12 June 2026. It then disabled Fable 5 and Mythos 5 for all customers. Anthropic disputed the directive. It said the government had given it only verbal evidence of a potential narrow jailbreak.
The Next Web, citing Reuters, reports the controls were lifted on 30 June. Access returned on 1 July, with usage capped at 50% of normal weekly limits until 7 July. The Next Web also reports that US Commerce Secretary Lutnick said no licence was needed, and that Anthropic agreed to risk detection, help with standards and reporting of malicious activity.
The practical lesson for a business is about supply. A model your workflow depends on can become unavailable at short notice.
Trends worth knowing
- Agents and coding. Epoch reports that coding-agent usage by OpenAI’s own researchers has recently been doubling roughly every month. That is a model fit to recent data, not a long-run rate.
- Computer use. Models that operate software on a screen are now a headline feature, as the OSWorld result above shows.
- MCP. Anthropic donated the Model Context Protocol to the Linux Foundation’s Agentic AI Foundation on 9 December 2025. Anthropic reports more than 10,000 active public servers and adoption by ChatGPT, Cursor, Gemini, Copilot and VS Code. A shared standard makes it easier to swap models without rebuilding your tools.
- Adoption. According to a Burges Salmon summary of the 2026 AI Index, organisational adoption of AI reached 88%.
EU AI Act delay
If you sell into Europe, note the change. The EU’s AI Omnibus was approved by the Council on 29 June 2026, according to Secure Privacy and datenrecht.ch. Addleshaw Goddard and Sidley report the new dates. Standalone high-risk obligations now apply from 2 December 2027. Product-embedded high-risk obligations apply from 2 August 2028. The changes take legal effect once published in the EU Official Journal. Secure Privacy reports that obligations for general-purpose AI models are unchanged. Transparency rules still apply from 2 August 2026, with a grace period to 2 December 2026 for watermarking on systems already on the market.
What this means for an Australian SME
- Do not lock into one model. Build so you can switch. The June suspension shows why.
- Start with the cheaper tier. Anthropic says Opus 5.5 matches Fable 5.1 on most work at lower cost. A small model like Haiku 5.5 may handle routine tasks.
- Test on your own tasks. Public benchmarks are mostly reported by the vendors themselves. Run five to ten real jobs from your business through two or three models and compare.
- Price your volume. Per-token prices differ enormously. Work out your monthly usage before you choose.
- Keep humans on risky steps. Payments, emails to customers and deletions should need approval.
If you want help choosing and testing models for your business, talk to us. We will run your real tasks and show you what the cheaper options can do.
Sources
- Anthropic newsroom, Anthropic, fetched 14 October 2026
- Anthropic lança Fable 5.1, Tecnoblog, 1-2 September 2026
- Project Glasswing, Anthropic, 7 April 2026 (updated 6 October 2026)
- Fable/Mythos access, Anthropic, 12 June 2026
- Anthropic Fable 5 export controls lifted, The Next Web, 1 July 2026
- GPT-6 Astra, OpenAI, undated, early September 2026
- OpenAI rolls out GPT-6 Astra model with cyber guardrails, Bloomberg Government, 3 September 2026
- Epoch AI Data Insights, Epoch AI, 16 September and 5 October 2026
- Google plans Gemini 4 release before year end, InfoWorld, 25 September 2026
- Meta hasn’t given up on open source: Muse Spark launches, NYU Shanghai RITS, 9 April 2026
- DeepSeek releases V4, NYU Shanghai RITS, 24 April 2026
- Qwen 3.5 medium series, NYU Shanghai RITS, 25 February 2026
- Donating the Model Context Protocol and establishing the Agentic AI Foundation, Anthropic, 9 December 2025
- AI in 2026: what does Stanford’s AI Index tell us?, Burges Salmon, 15 April 2026
- EU lawmakers reach provisional agreement to delay key EU AI Act obligations, Sidley, 22 June 2026
- EU AI Act: AI Omnibus formally adopted, Addleshaw Goddard, 1 July 2026
- EU AI Act Digital Omnibus: The New High-Risk AI Deadlines After Council Approval, Secure Privacy, 10 July 2026
- Digital Omnibus on AI Also Passed by the Council, datenrecht.ch, 1 July 2026