OpenAI and Anthropic Slash LLM Prices by Up to 80% Amid Intensifying Competition

OpenAI and Anthropic Slash LLM Prices by Up to 80% Amid Intensifying Competition

Lead – Since June 2026, OpenAI and Anthropic have repeatedly cut the per‑token pricing of their large language models (LLMs), with reductions as steep as 80 %. The price drops, reported on Hacker News, affect both input and output token rates as well as cache‑read fees, and appear to be driven by a rapid escalation of competition from other AI providers, including open‑weight models such as DeepSeek’s Flash series.

Timeline of Model Announcements and Price Changes

| Date | Provider | Model | Input $/MTok | Output $/MTok | Cache‑read $/MTok | |------|----------|-------|--------------|---------------|-------------------| | 26 Jun 2026 | OpenAI | GPT‑5.6 Sol | $5 | $30 | $0.50 | | 26 Jun 2026 | OpenAI | GPT‑5.6 Luna (small‑class) | $1 | $6 | $0.10 | | 30 Jun 2026 | Anthropic | Claude Sonnet 5 | $2 | $10 | $0.20 (temporary) | | 24 Jul 2026 | Anthropic | Claude Opus 5 | $5 | $25 | $0.50 | | 30 Jul 2026 | OpenAI | GPT‑5.6 Luna price cut ≈ 80 % → $0.20 / $1.20 / $0.02 | | 10 Aug 2026 | Anthropic | Sonnet 5 price cut made permanent | | 21 Aug 2026 | OpenAI | GPT‑5.6 Sol price cut ≈ 20 % → $4 / $20 / $0.40 (3‑month term) | | 01 Sep 2026 | Anthropic | Fable 5.1 | $10 | $50 | $0.25 (75 % cache‑read reduction) | | 03 Sep 2026 | OpenAI | GPT‑6 Astra | $10 | $50 | $1 | | 22 Sep 2026 | OpenAI | GPT‑6 Sol | $2 | $10 | $0.20 (½ input/output of discounted GPT‑5.6 Sol) | | 22 Sep 2026 | OpenAI | GPT‑6 Luna | $0.10 | $0.50 | $0.01 | | 22 Sep 2026 | Anthropic | Claude Opus 5.5 | $4 | $20 | $0.20 (20 % I/O cut, larger cache cut) | | 29 Sep 2026 | OpenAI | GPT‑6.1 Sol | $2 | $10 | $0.10 (cache‑read cut ½) |

According to Hacker News, the cumulative effect of these adjustments is a 75 % reduction in the total cost of using the GPT‑6.1 Sol family compared with the original GPT‑5.6 Sol rates announced in June.

Cache‑Read Costs and Their Impact

The source notes that, when LLMs are used as agents for coding, cache‑read fees account for 96 %–99 % of total token expenditure because the cache stores the full conversation history for each turn. With cache hit rates typically between 90 % and 100 %, a 75 % reduction in cache‑read cost—as seen with Anthropic’s Fable 5.1—translates into an overall cost decline of roughly 70 %–75 % for agentic workloads.

Open‑Weight Competitors Pressuring Prices

According to Hacker News, the price cuts coincide with the rise of open‑weight models that offer lower advertised rates. DeepSeek’s Flash V4 series, released in April 2026 at $0.14 / $0.28 / $0.0028, quickly became popular because its cache‑read proportion of 2 % is markedly lower than the 5 %–10 % typical of OpenAI and Anthropic models. The updated Flash V4.1 1017 variant, listed on OpenRouter, now costs $0.15 / $0.60 / $0.003 and delivers a benchmark score of 45 on the Artificial Analysis index, up from 24 for the original V4.

OpenRouter aggregates multiple inference providers. As of 24 Oct 2026, 29 distinct providers host the Flash V4.1 1017 model, compared with 13 providers for the original V4. The platform allows developers to request the “lowest price” endpoint, effectively exposing end‑users to the cheapest cache‑read rates across the ecosystem.

Competitive Dynamics Between OpenAI and Anthropic

The timeline illustrates a clear tit‑for‑tat pattern: each company’s price reductions closely follow the other’s announcements. According to Hacker News, OpenAI’s 80 % cut to GPT‑5.6 Luna in July appears to target DeepSeek’s low‑cost offering, while Anthropic’s permanent reduction of Sonnet 5’s launch price in August mirrors OpenAI’s earlier discount.

Anthropic’s September 22 release of Opus 5.5 explicitly states that the model “requires less compute to serve than Opus 5, and its pricing reflects that,” indicating a deliberate effort to lower both input/output and cache‑read fees. OpenAI’s subsequent launch of GPT‑6.1 Sol with a halved cache‑read cost further tightens the pricing battle.

Market Implications

According to Hacker News, the aggregate spend on LLMs continues to rise despite falling per‑token costs, because model capabilities are improving faster than prices are declining. Vercel’s AI model gateway data show a 43.3 % reduction in cost‑per‑token relative to April 2026, yet overall usage has expanded.

Open‑weight models such as DeepSeek and the newly announced Moonshot AI Kimi K3.1 Pro (which matches Opus 5.5’s benchmark score of 58 at half the input/output price) are eroding the moat that larger providers once enjoyed. As reported by Hacker News, the only remaining differentiator for OpenAI and Anthropic is higher‑quality performance, prompting continued multi‑billion‑dollar R&D investments.

Brief Analysis (Opinion)

The rapid succession of price cuts suggests that both OpenAI and Anthropic view token pricing as a strategic lever rather than a static cost structure. By slashing fees, they aim to retain enterprise and developer demand that might otherwise shift to cheaper open‑weight alternatives. The focus on reducing cache‑read costs—where the bulk of expense resides for agentic tasks—demonstrates a nuanced understanding of where savings can be achieved without compromising model quality.

Conclusion

From June 2026 to late September 2026, OpenAI and Anthropic have collectively reduced the cost of their flagship LLMs by up to 80 %, a shift driven by intense competition from open‑weight models and the need to keep agentic workloads affordable. While per‑token prices have plummeted, the overall market spend continues to climb as developers adopt increasingly capable models. The next few months will likely reveal whether price alone can sustain market share or if performance breakthroughs will become the decisive factor.

Frequently Asked Questions

What prompted OpenAI’s 80 % price cut for GPT‑5.6 Luna?**

According to Hacker News, the cut was likely intended to compete directly with low‑cost open‑weight models such as DeepSeek’s Flash V4, which offered significantly cheaper cache‑read rates. **Q: How do cache‑read costs affect the total expense of using LLMs for coding agents?** A: As reported by Hacker News, cache‑read fees can represent 96 %–99 % of token expenditure in agentic coding scenarios, so a 75 % reduction in cache‑read pricing can lower overall costs by roughly 70 %–75 %. **Q: Are the recent price reductions accompanied by a decline in model quality?** A: According to Hacker News, the price cuts have not been accompanied by observable quality degradation; in several cases, newer models (e.g., GPT‑6.1 Sol) match or exceed the performance of earlier, higher‑priced versions. Source: Hacker News

Source: Hacker News — Read more

💻 Related Articles