LLM Price Watch Weekly Digest
No price changes this week across the five providers we track — Claude, GPT, Gemini, DeepSeek, and Grok all held steady. But there was a quiet flurry of new model IDs showing up on OpenRouter, all of them batch variants, and worth knowing about if you're running async workloads.
Here's what we spotted on August 29th.
DeepSeek: Two New Batch Models
DeepSeek added two new batch-mode models:
deepseek/deepseek-v4-pro-0813:batchdeepseek/deepseek-v4-flash-0731:batch
The naming convention tells you a few things. The -pro vs -flash split mirrors what we've seen before — pro for heavier reasoning tasks, flash for faster and cheaper throughput. The date suffixes (0813, 0731) suggest these are specific checkpoints, which is useful if you care about reproducibility or are comparing outputs across versions.
Batch mode generally means you submit a job and get results back asynchronously, usually at a lower cost per token than the real-time endpoint. If you're doing large-scale data processing, eval runs, or document analysis where latency doesn't matter, batch is almost always the right call. Worth checking whether DeepSeek's batch pricing lands meaningfully below their synchronous rates.
Google: Gemma 4 31B Batch
google/gemma-4-31b-it:batch
Gemma 4 at 31B is a reasonable mid-size open-weights model. The it tag means instruction-tuned. Seeing it show up as a batch endpoint on OpenRouter suggests Google is continuing to expand the Gemma family's API availability beyond just the hosted Gemini models. For teams that want Google's open-weights lineage without committing to Gemini pricing, this is worth a look — especially for batch jobs where you can afford to wait.
OpenAI: Two "OSS" Models
These two caught my eye:
openai/gpt-oss-120b:batchopenai/gpt-oss-20b:batch
The oss label is interesting. OpenAI hasn't broadly publicized models under that naming scheme, so I'm not going to speculate too much about what's behind it. What's observable is that there's a 120B and a 20B variant, both batch-only for now, both appearing on OpenRouter. The size gap between them is large — 120B puts it in flagship territory, 20B is more of an efficient workhorse. If these are genuinely open or open-weight models from OpenAI, that would be notable, but I'd wait for official documentation before drawing conclusions.
For now, treat them as "spotted in the wild" and monitor whether pricing and model cards show up.
What This Week Means Practically
If you're actively choosing models right now:
Batch is underused. Most developers I talk to default to synchronous endpoints even when their use case has no real latency requirement. If you're running nightly pipelines, processing large document sets, or doing offline evals, switching to batch can cut costs significantly. This week's additions give you more batch options across providers.
DeepSeek continues to be prolific. Two new models in one week, both with dated checkpoints, suggests active development. If you're relying on DeepSeek for production workloads, pinning to a specific checkpoint (like 0813) is probably smarter than using a floating alias.
OpenAI's OSS naming is worth watching. No firm conclusions yet, but if these turn out to be open-weight models with competitive pricing, that changes the calculus for teams currently choosing between self-hosting and API access.
No pricing drama this week — just new surface area to explore. I'll flag it here if anything changes on the cost side.
We track live pricing and new model appearances for Claude, GPT, Gemini, DeepSeek, and Grok at llmpricewatch.com.
Top comments (0)