Dor Josef
← All writing

Opus 5.5 Cuts Prices, GPT-5.6 Sol Drops 50%, and Qwen 3.8 27B Wins the Local Crowd

Opus 5.5's cheaper output tokens and GPT-5.6 Sol's 50% price cut collide with Qwen 3.8 27B's breakout on Hacker News, while DeepSeek's plugin-first Harness framework and a maturing 27B-class local model category show where the agent era is heading.

· Updated 2026-09-30

  • AI
  • Local-Models
  • Price-war
A playful AI marketplace where three competing AI systems are dramatically lowering their price tags, while a compact local AI machine on a developer's desk becomes powerful enough to compete with giant cloud systems.

The price of intelligence is falling - from both ends

Anthropic's Opus 5.5 is the week's biggest model story. It's the company's first frontier release since Amodei's "pace the frontier" slowdown - and it beats the bigger Fable model on coding and benchmarks while cutting output tokens from $25 to $20 per million tokens.

OpenAI is matching the push. GPT-6 Luna and Sol landed this week, joining GPT-Live-1 in a rapid release cadence - 10 models in 6 months - and on Hacker News, reports that GPT-5.6 Sol's price was cut 50% (per OpenRouter) on OpenAI's top vision model changed the local-vs-API math.

The open-weights verdict is more mixed. Simon Willison's review of Qwen 3.8 27B - "excellent, but it defaults to overthinking things" - was the talk of HN, corroborated by the model's 52 score on the Artificial Analysis leaderboard. Practitioner mood: small local models are genuinely good now, but they carry an "overthinking" tax.

The 27B sweet spot is a product category now

HuggingFace's trending list tells the same story from the local side. DeepSeek's 763B MoE DeepSeek-V4.1-Flash was trending #1 by downloads - server-only, since 128GB can't hold even an aggressive MoE offload of this size - but the actionable item was DavidAU's Qwen3.6-27B IMatrix GGUF: a 27B quant tuned for coding and agentic work that runs in 128GB unified memory with room for long context at Q8, ready to drop into an Ollama slot tonight.

Two more data points from the same board: DeepSeek-V4-Pro - 1.6T total params, 49B active, 1M context, MIT-licensed - now appears in community datasets, with quants to watch; and Gemma 4 31B is in the Dell Enterprise Hub with MTP speculative-decoding variants for the same hardware class.

GitHub trending confirms the shift from model quality to orchestration and token efficiency. DeepSeek's Harness agent framework - where every tool, subagent, and memory is a swappable plugin - hit 163K stars in ~41 days, a major signal that DeepSeek is shipping an agent platform. ByteDance's Deer Flow ships a long-horizon "SuperAgent" that researches, codes, and creates in sandboxes, while Headroom claims 60–95% token reduction on compressed JSON tool output before it hits the LLM.

The agent PR question

The human side of the agent wave was the day's sharpest HN discussion: a developer used an AI agent to fix an open-source bug, and someone asked to ban them. The community is actively debating whether AI PRs are welcome at all - a live question for any team shipping agentic workflows, and part of why the agent-harness war is on, with DeepSeek, ByteDance, and independent teams all shipping competing "agent runtime" frameworks.

Signal to watch

Watch the convergence: Opus 5.5's cheaper tokens, GPT-5.6 Sol's price cut, and 27B-class open models that are genuinely good at coding are all compressing the cost of intelligence from both ends. If Anthropic's Sonnet 5.5 / Haiku 5.5 "coming weeks" land at the new price points, the local-vs-API math flips for many teams.

Sources