Anthropic and OpenAI slash token prices with Claude Opus 5.5 and GPT‑6 Sol/Luna

Anthropic unveiled Claude Opus 5.5, then OpenAI followed with GPT‑6 Sol and GPT‑6 Luna, both slashing token costs by half and introducing new pricing tiers that could reshape the AI‑as‑a‑service market.

Key points

• Claude Opus 5.5 drops input price to $4/million, output to $20/million.
• GPT‑6 Sol and GPT‑6 Luna are priced at $0.10/$0.50 and $0.10/$0.50 per token, half of GPT‑5.6.
• Cache read prices fell 60 % in Opus 5.5.
• Grok 4.7 now matches GPT‑6 Sol on input pricing.
• Opus 5.5’s max output token limit is 128,000, matching other Claude models.

What happened

On September 22, Anthropic released Claude Opus 5.5, followed an hour later by OpenAI’s GPT‑6 Sol and GPT‑6 Luna. The new models introduced significant price cuts: Opus 5.5 lowered input and output costs by 20 % from its predecessors, while GPT‑6 Sol and GPT‑6 Luna cut their token rates to $0.10 for input and $0.50 for output, effectively halving the promotional rates of GPT‑5.6. The report notes that GPT‑5.6 is scheduled for a 25 % price increase in November, making GPT‑6’s pricing even more attractive.

The source also documents a 60 % drop in cache read pricing for Opus 5.5, a move that benefits long‑running agentic conversations where cached tokens dominate. Anthropic announced upcoming Sonnet 5.5 and Haiku 5.5, while OpenAI’s new models already undercut Haiku 4.5’s $1/$5 rates.

The backstory

Anthropic’s Opus line has evolved from Opus 4.5–5.0, all priced at $5/million input and $25/million output. The incremental upgrades focused on communication clarity and token efficiency, but critics cited slow reasoning. OpenAI’s GPT‑5 series introduced higher‑tier models like GPT‑5.6 Luna, praised for performance and affordability, before the September releases.

Price war dynamics

The aggressive price cuts from both vendors signal a new phase of competition. By offering GPT‑6 Sol and Luna at $0.10/$0.50, OpenAI undercuts Grok 4.7’s $2/$6 rates and matches Anthropic’s own Opus 5.5 level. This forces developers to re‑evaluate cost per token against performance, especially for high‑volume use cases.

Technical trade‑offs and limitations

Opus 5.5’s 128,000‑token output cap and a 60 % cache read reduction improve efficiency but may limit very long documents. The source shows that Opus 5.5 can stall during complex reasoning, as evidenced by a failed SVG pelican test where the model hit its token ceiling mid‑reasoning. GPT‑6 Sol and Luna, while cheaper, may offer similar or slightly lower throughput compared to GPT‑5.6 Luna, which was previously the top performer.

What builders and businesses should do

• Recalculate token budgets for existing GPT‑5.6 workloads to assess savings.
• Test GPT‑6 Sol and Luna in production scenarios to benchmark latency and accuracy.
• Consider cache‑read pricing when designing conversational agents.
• Monitor Anthropic’s upcoming Sonnet 5.5 and Haiku 5.5 for mid‑tier options.
• Evaluate whether Opus 5.5’s 128,000‑token limit suffices for your longest documents.

The bottom line

Watch OpenAI’s pricing for the next tier of models in early October, and keep an eye on Anthropic’s Haiku rollout. The cost advantage may shift adoption curves for mid‑tier AI services.