The dialog within the AI {industry} has largely been centered on slowing down AI improvement or “pacing the frontier,” however that apparently will not cease Anthropic and OpenAI from releasing new fashions. Each firms are iterating on their earlier releases, Fable 5.1 and GPT-6 Astra, with new fashions that supply related ranges of efficiency however at decrease prices.
Within the case of Claude, Anthropic says its new Opus 5.5 mannequin is healthier at dealing with complicated work, “discovering and fixing inefficiencies in software program” and monetary evaluation and enterprise work — all pitches laser-targeted at Anthropic’s enterprise clients. Anthropic’s benchmarks declare Opus 5.5 additionally scored higher at agentic coding than GPT-6 Astra on each Terminal-Bench 4.0 and FrontierCode v1.1 (Foremost), which may make it extra interesting for builders.
Enhancements prolong exterior of the duties the mannequin can deal with and to the way it works: Anthropic claims Opus 5.5 produces writing that is simpler to grasp, and that it “tried to bypass boundaries round 85 % much less usually” than previous fashions, suggesting it’s going to disobey instructions much less usually. Equally essential, the corporate claims it prices much less per token to make use of ($4 for enter tokens and $20 for output tokens, in comparison with the $5 enter and $25 output of Opus 5).
OpenAI’s new GPT-6 Sol and GPT-6 Luna fashions are equally centered on effectivity. The brand new fashions had been educated utilizing related strategies to GPT-6 Astra and supply enhancements throughout skilled work, coding and laptop use, however for as much as 50 % cheaper than the promotional pricing of their GPT-5.6 counterparts. Particularly, GPT-6 Sol enter tokens price $2, whereas output tokens price $10. For Luna, enter tokens price $0.10, whereas output tokens price $0.50.
When it comes to enhancements, each new GPT-6 fashions are higher at getting details proper, with Sol making about half as many errors as its predecessor, based on OpenAI. In coding, the corporate says GPT-6 Sol is ready to match the efficiency of Fable 5.1 at a decrease price, too. The brand new fashions additionally talk extra clearly, and due to enhancements to immediate caching, OpenAI says they’ll reuse extra context and reply quicker.
OpenAI says the GPT-6 fashions are extra aligned than earlier than primarily based on the corporate’s homegrown checks. For instance, they lie much less concerning the outcomes of their coding work and deny much more makes an attempt to bypass safeguards after they obtain an unsafe command. Like Anthropic, OpenAI has additionally proposed its personal standards for third-party evaluators to make use of when judging the progress and security of AI improvement. Whether or not these proposals will result in some type of industry-wide normal stays to be seen although.
Anthropic’s Opus 5.5 mannequin is out there to builders now via Claude, Amazon Net Providers, Google Cloud and Microsoft Azure. OpenAI’s GPT-6 Sol and GPT-6 Luna are each accessible in ChatGPT Work and Codex for Plus, Professional, Enterprise and Enterprise clients. Free customers and Go subscribers could have entry to solely GPT-6 Luna within the firm’s desktop app.
