Anthropic's best model struggles to win users as cheaper tools thrive
An FT report puts Anthropic's annualized revenue at $65B in July, up from $47B in May, with the company still projecting a profitable Q3 and over 6,000 customers spending $100K+/year — strong top-line numbers. But billing data from Ramp across 70,000 companies tells a different story about which models people actually reach for: Opus 4.8 still holds 28% of coding spend, while the pricier Fable 5 sits at just 8% and the newer Opus 5 (released July 24) captured only 3.5% by month's end. OpenAI's own revenue jumped 35% to over $40B after GPT-5.6's July launch. Read alongside this week's continuing Sol/Luna price cuts, it's a picture of frontier labs racing on price even as headline revenue keeps climbing — premium capability isn't where the volume is.
Sources & depth
I gave Qwen 3.8 27B a reverse-engineering job and it finished in 30 minutes
A hands-on test pointed Qwen3.8-27B — a 27B dense model that fits in 17GB VRAM — at a task normally reserved for frontier cloud models: reverse-engineering a commercial app's license-verification scheme it had never seen in training. Running locally on a Grace Blackwell mini workstation (SGLang, NVFP4 quantization, DFlash2 speculative decoding) using only static analysis and standard Bash tools inside the Pi harness, the model disassembled ARM64 code, recovered a deliberately obscured RSA public key, and produced a working authentication-bypass proof-of-concept in 30 minutes — including catching and correcting its own failed key-reconstruction attempt without being prompted. It's the latest proof point for the Qwen3.8 family's momentum since its day-1 GGUF release (Unsloth/Ollama support, alongside the 2.4T/95B-active FP8 Max variant), which the desk's already-open qwen38-max-battery pitch is tracking toward a weekly shortlist call.
Sources & depth
Quoting Drew Breunig
Quoted by Simon Willison, developer Drew Breunig argues Fable reset the calculus for coding-harness investment: before Fable, spending time optimizing an agent's harness or context strategy felt wasteful, since a cheaper, better model would arrive soon and paper over the problem anyway. Fable itself is excellent but priced high enough that Opus, GPT-5.6, K3, and GLM remain "good enough" for most coding work — so instead of waiting for raw capability to solve everything, teams are now deliberately deciding which model handles which task by cost. It's a sharp read on where LLM-assisted coding economics are heading as the price gap between frontier and "good enough" models widens — a smaller companion data point to the Anthropic pricing story above.
Sources & depth