New AI Models Released: Opus 4.6 and GPT-5.3-Codex Redefine Specialization

Anthropic and OpenAI have recently launched their flagship models, Opus 4.6 and GPT-5.3-Codex, respectively, with a mere 27-minute difference in their launch times. Both models boast impressive benchmark leads, albeit in different areas. Opus 4.6 excels in reasoning tasks such as Humanity’s Last Exam, GDPval-AA, and BrowseComp, while GPT-5.3-Codex takes the lead in coding tasks like Terminal-Bench 2.0.

The pricing of these models, however, raises eyebrows. Opus 4.6 is significantly more expensive than its counterparts, with a price spread that is difficult to ignore. For instance, the cost of input and output for Opus 4.6 is $5.00 and $25.00, respectively, whereas Gemini 2.5 Pro costs $1.25 and $10.00. Open-source alternatives are available at a fraction of the cost, with some options being 50x less expensive.

The benchmark gap between these models is expected to justify the price difference, but for many tasks, it does not. The inclusion of 1M context is becoming a standard feature, with Opus 4.6 offering 1M tokens (beta, 2x pricing past 200K) and Gemini already providing 1M at standard pricing. The real differentiator lies in the retrieval quality at this scale, with Opus 4.6 scoring 76% on MRCR v2 (8-needle, 1M), the strongest result so far.

The market reaction to these launches was immediate, with Thomson Reuters stock falling 15.83% and LegalZoom dropping nearly 20%. The tradeoff between different model capabilities is becoming increasingly apparent, with Opus 4.6 receiving complaints about writing quality from early users. The theory is that RL optimizations for reasoning may have degraded prose output, highlighting the challenge of creating a single model that excels across all tasks.

As the AI landscape continues to evolve, it is clear that the frontier is fragmenting by task type, with no single model emerging as a clear winner. For more information and full benchmarks, please refer to the source article: Claude Opus 4.6: 1M Context, Agent Teams, Adaptive Thinking, and a Showdown with GPT-5.3

Photo by lebih dari ini on Pexels
Photos provided by Pexels