OpenAI Cuts GPT-5.6 Prices; DeepSeek Fires Back

OpenAI slashed API prices for two of its GPT-5.6 models on July 30. Price tags for Luna, its smallest model, fell 80% to $0.20 per million input tokens and $1.20 per million output tokens, while the mid-tier Terra dropped 20% to $2 and $12. The flagship Sol kept its pricing at $5 and $30.

The reductions come as enterprise clients grow increasingly frustrated with rising token bills while low-cost Chinese models gain market share. A price cut OpenAI was reported to be considering in June has materialized in just seven weeks.

July Sales Outpaced the Entire Second Quarter

The confidence behind these price cuts stems from record monthly performance. OpenAI launched the GPT-5.6 series—comprising Sol, Terra, and Luna—on July 9, and recently informed employees that July's annualized revenue outpaced the company's entire second-quarter revenue.

CFO Sarah Friar and board chair Bret Taylor credited GPT-5.6, ChatGPT Work, and the Codex coding agent for the growth. "You're seeing people who went deep on Claude Code, ended up with a very high bill, and started looking for an alternative," Taylor told staff.

The target is the budget-friendly coding agent market. Following the cuts, Luna's combined pricing of $1.40 per million tokens undercuts Google's Gemini 3.5 Flash-Lite at $2.80. A report by CNBC framed the move as a direct response to enterprises tightening their AI budgets. Cutting prices in the middle of a hot streak reads as offense, not defense.

DeepSeek V4 Flash Closes In on Claude Opus 4.8

DeepSeek responded within a day on the other side of the budget market. On July 31, the company released the official V4 Flash build (designated 0731). The model retains the 284-billion-parameter architecture of April's preview but has been re-post-trained, and the update was seamlessly applied to existing API calls.

The published benchmark results drew immediate attention. The new build scored 82.7 on Terminal-Bench 2.1, trailing Anthropic's previous flagship Claude Opus 4.8 (85.0) by only 2.3 points. On the GitHub issue-resolution benchmark DeepSWE, V4 Flash posted 54.4 compared to 58.0 for Opus 4.8. Crucially, it outperformed DeepSeek's own V4 Pro preview across all nine published agent benchmarks.

At a combined rate of $0.42 per million tokens, the model costs less than one-third of Luna's new pricing. DeepSeek announced that the official V4 Pro release will follow soon. With Moonshot AI's Kimi K3 and other Chinese models launching, the coding agent market has entered an intense race on both capability and unit cost.

Anthropic Squeezed by Tokenizer and Price War

This competitive landscape places Anthropic in a difficult position. Claude remains the most expensive model family on the market. Furthermore, a new tokenizer introduced in late June counts the same TypeScript file as containing up to 73% more tokens than OpenAI's GPT-5.x tokenizer, according to an analysis reported by The Register. Anthropic has acknowledged that the updated tokenizer can map the same input to roughly 1.0 to 1.35 times more tokens. The tokenizer change alone amounts to an effective price hike without any change to official list prices.

Even the performance lead is wobbling. On the independent SWE-bench Verified leaderboard, Claude Opus 5, released on July 24, holds first place at 97.0% — but GPT-5.6 Sol sits at 96.2%, effectively level, while flagship Fable 5 trails at 95.0%. What remains of the gap is the harder SWE-bench Pro, where Fable 5's 80.3% clearly beats Sol's 64.6%.

The problem is what that remaining edge costs. Opus 5 runs a combined $30 per million tokens — more than 20 times Luna — and Fable 5's $60 sits over 70% above Sol's $35. In the Luna and Flash price band, Anthropic fields no Claude model at all.

Surrounded by rivals selling near-equal performance for far less, Anthropic faces a choice between matching the cuts and rebuilding a clear lead. With OpenAI lowering rates and DeepSeek raising performance in the same week, the next move in the price war belongs to Anthropic.