Tech • AI • Robotics • Game
Opus 5.5, GPT 6 Sol, Mimo 2.6, now that’s a slowdown!
9.4/10TL;DR
Anthropic has raised the competitive pressure in AI by launching Claude Opus 5.5, a stronger and cheaper flagship model, while OpenAI, xAI, and Xiaomi escalate a fast-moving price and performance battle across proprietary and open models.
KEY POINTS
Anthropic launches Claude Opus 5.5
Claude Opus 5.5 is positioned as the first model in a new 5.5 family and is being presented as a major upgrade over Opus 5. The model is advertised as delivering performance at least on par with Claude Fable 5.1, and on several benchmarks it reportedly surpasses that earlier top-tier model. Anthropic is also preparing a broader lineup around the same generation, with smaller variants expected later.
A 40% drop in operating cost
Anthropic says Opus 5.5 costs 40% less to run than Opus 5, combining lower token pricing with greater efficiency per task. Pricing moves from $5 per million input tokens and $25 per million output tokens on the previous Opus generation to $4 and $20 respectively. The company is also arguing that lower token consumption reduces total task cost, not just headline API rates.
Benchmarks suggest a widening lead
On coding, agentic use, reasoning, knowledge tasks and computer-use benchmarks, Opus 5.5 is presented as leading the field. In the Artificial Analysis Intelligence Index, it is described as opening a noticeable gap over competing frontier models. That would mark a significant shift in a market where leading systems had recently clustered much more closely together.
Focus on developers and long tasks
Anthropic is targeting software development, migrations and long-context work as prime use cases. In one cited programming-language translation test, a task that previously took 12 hours with Fable 5.1 was completed successfully by the new model at about 51% lower cost. The positioning is clear: Opus 5.5 is being pitched as a production model for developers and high-end automation.
Anthropic also pushes style and safety
Beyond raw capability, Anthropic claims the model writes in a more natural style and follows formatting instructions more closely, addressing complaints about overly formulaic outputs. On safety, the company says reasoning is easier to inspect, alignment is handled during training, zero data retention remains in place, and watermarking is used to comply with European Union rules. Some high-risk cybersecurity or biological prompts are routed or blocked more aggressively.
OpenAI expands GPT-6 with Sol and Luna
OpenAI released GPT-6 Sol and GPT-6 Luna on the same day, extending the GPT-6 range below GPT-6 Astra. Sol is aimed at repeated complex work such as coding, debugging and data analysis, while Luna targets high-volume, lower-cost tasks including summaries, extraction and short responses. The strategy appears centered on segmentation and affordability rather than a leap in top-end performance.
Lower OpenAI prices, weaker benchmark story
OpenAI has sharply reduced costs, especially for lighter workloads and cached prompts, with caching priced at roughly 10% of the standard rate. But benchmark comparisons appear less flattering: Sol and Luna are described as trailing some earlier or rival models on pure performance. That makes the OpenAI pitch more about cost per completed task than benchmark leadership.
Xiaomi enters with open model Mimo 2.6
Xiaomi has emerged as a serious challenger in open AI with Mimo 2.6, an agent-focused model aimed at coding, tool use and autonomous action. It comes in Pro and Flash variants, with Xiaomi openly acknowledging that it still trails the strongest American frontier models but claims to be closing the gap fast. In several cost-versus-capability comparisons, Mimo 2.6 Pro is presented as especially competitive.
Open release and unusually low training costs
Xiaomi released Mimo 2.6 under an MIT license with technical reports, training code and documentation, making it one of the most permissive major model launches to date. The company says reinforcement learning lasted just 6 days, with training costs of about $850,000 for Flash and $2.6 million for Pro. That cost disclosure sends a strategic signal in the broader China-United States AI race.
xAI cuts price with Grok 4.7 but stumbles
Grok 4.7 arrived just ahead of the larger launches and is markedly cheaper than leading rivals, roughly around half the price of top models from OpenAI, which themselves sit below Anthropic’s premium tier. But the model is described as less reliable than expected and still expensive in real task terms because it uses more tokens. For coding and developer workflows, that weakens its value despite the headline discount.
Amazon blocks Meta’s shopping agent
In a related sign of how AI agents are colliding with platform business models, Amazon has blocked Meta’s shopping agent from accessing its marketplace. The dispute raises a larger commercial issue: autonomous agents may bypass sponsored placement and ad-driven discovery systems that generate billions in revenue. The clash suggests that as AI agents become more capable, access to major digital platforms may become a central battlefield.
CONCLUSION
The AI market is shifting from a race over raw intelligence alone to a three-way contest over performance, real task cost, and platform control. For now, Anthropic appears to have seized the performance lead, while OpenAI, xAI, and Xiaomi are forcing the next phase of competition on price and distribution.
Ask a question