Teknium says Anthropic still leads on long-context coherence and cost
Teknium · x · 2026-07-22
Teknium says long-context quality and affordability remain the main bottleneck for practical model use.
After testing GPT-5.6 Sol and Terra, the post argues that Anthropic is still the clear leader in long-context coherence and may even be cheaper at that scale. It claims OpenAI charges about 2x once prompts exceed 350K tokens, but the models are not worth using there because coherence falls apart. By contrast, the author says Opus and Fable stay coherent even at 800K+ context.
The post also wonders what happened to Magic.dev’s promised 100M-token context window, suggesting that the idea may not have worked out in practice.
More from Models
- Google Reportedly Releasing Three New Gemini Models Including 3.6 Flash — xiaohu · 2026-07-22
- Trick Discovered: Typing Arabic in English Bypasses Limits in o3 — willdepue · 2026-07-22
- Users are switching GPT-5.6 variants to dodge cybersecurity request blocks — ivan_bezdomny · 2026-07-22
- A discussion of post-training incentives and long-horizon instruction following — xuanalogue · 2026-07-22
- A frontier AI control stack proposes logs, scans, defenses, and breach plans — sjgadler · 2026-07-22
- Fireworks says Kimi K3 handles 72–96% of agent traffic at up to 50x lower cost — eliebakouch · 2026-07-22