Every week I pull the actual price data from our tracker at LLM Price Watch and report what changed. This week there's a real price drop worth knowing about, plus a flood of new models that showed up on OpenRouter overnight.
The Price Change: DeepSeek V3
DeepSeek V3 dropped again. As of September 24th, it's sitting at $0.27 per million input tokens and $0.41 per million output tokens.
To put that in practical terms: if you're running a mid-sized RAG pipeline pushing a few hundred million tokens a month, you're looking at real money saved versus most of the alternatives. DeepSeek has been on a consistent downward trajectory on pricing, and V3 in particular has become a go-to recommendation for teams that need solid general-purpose performance without paying frontier model prices.
For comparison, if you were previously paying $0.38/M input (roughly where V3 was a few months back), the new rate is a meaningful cut. Output-heavy workloads — think long-form generation, detailed summarization, agentic loops — benefit most here since output tokens are typically where costs pile up.
If you haven't benchmarked DeepSeek V3 for your use case recently, it's worth a look now.
New Models: OpenAI's GPT-6 Family
Eight new OpenAI model IDs appeared on OpenRouter on September 23rd:
gpt-6-luna-progpt-6-luna-pro:batchgpt-6-lunagpt-6-luna:batchgpt-6-sol-progpt-6-sol-pro:batchgpt-6-solgpt-6-sol:batch
Our tracker spotted all eight at the same timestamp, which suggests a coordinated rollout rather than a staged one. The naming pattern — Luna and Sol as sub-variants, with Pro tiers and Batch variants for each — looks like OpenAI is following the same tiering playbook they used with the GPT-4o family.
I don't have pricing data for these yet (they appeared as new model IDs without confirmed per-token rates in our snapshot), so I'd treat any numbers you see floating around as unverified until they stabilize. Worth watching.
New Models: Anthropic Claude Opus 5.5
Two more new IDs from Anthropic:
claude-opus-5.5claude-opus-5.5:batch
Opus has historically been Anthropic's most capable and most expensive tier, so a point-release here is interesting. Again, no confirmed pricing in our snapshot yet — our tracker caught the model IDs before rate data was available.
New Models: DeepSeek V4.1 Flash Batch and Grok 4.7
DeepSeek also snuck in a batch variant of what appears to be a V4.1 Flash model (deepseek-v4.1-flash:batch), first seen September 23rd. No pricing confirmed yet, but the "flash" naming convention from DeepSeek has previously meant a smaller, faster, cheaper variant — so this could be interesting for high-volume, latency-tolerant workloads.
And on the xAI side, grok-4.7 appeared on September 22nd — a day ahead of the others. No pricing data in our snapshot for this one either.
What to Actually Do With This
Here's the practical summary:
- If you're using DeepSeek V3, check your current rate against $0.27/$0.41 — you may already be on the new pricing depending on your provider, but it's worth verifying.
- The GPT-6 and Claude Opus 5.5 launches are real, but hold off on budgeting around them until pricing stabilizes and benchmarks are out.
- DeepSeek V4.1 Flash batch and Grok 4.7 are names to file away. Watch for pricing.
I'll update as confirmed rates come in. The tracker runs daily checks, so if anything shifts mid-week you can catch it live.
Full live pricing for Claude, GPT, Gemini, DeepSeek, and Grok is at llmpricewatch.com.
Top comments (0)