DEV Community: Anders Rasmussen

DeepSeek Flash Pricing Swung Twice This Week, and GPT-6.1 Terra Just Showed Up

Anders Rasmussen — Sun, 01 Nov 2026 10:01:37 +0000

Here's what our price tracker picked up over the last seven days across Claude, GPT, Gemini, DeepSeek, and Grok.

DeepSeek V4.1 Flash changed price twice in five days

This is the headline for practical reasons. If you're using DeepSeek V4.1 Flash in production, the input price moved twice in one week:

Oct 26: $0.02 input / $1.00 output per million tokens
Oct 31: $0.045 input / $0.50 output per million tokens

So input got more than doubled (from $0.02 to $0.045), while output got cut in half (from $1.00 to $0.50). The net effect depends entirely on your workload. If you're doing something input-heavy — long context retrieval, document processing, big system prompts — this week ended up more expensive for you. If you're generating a lot of output relative to input, you came out ahead.

The bigger point is that two price changes in five days on the same model is unusual. Most providers move prices slowly and announce it. Seeing it shift mid-week without fanfare is a good reminder that relying on cached cost estimates in your head (or in a spreadsheet from last month) can quietly break your budget assumptions. Worth re-checking before you commit to a contract or a long-running batch job.

Kimi K3.1 Pro appeared at $4 input / $20 output

We also spotted a pricing record for Kimi K3.1 Pro on Oct 27: $4.00 per million input tokens, $20.00 per million output tokens. That puts it in the same general tier as GPT-4-class models — not cheap, not the most expensive thing out there, but clearly positioned as a premium offering rather than a budget option.

We don't have a prior price snapshot for this model to compare against, so I can't tell you if this is a change or just the first time we caught it. Either way, if Kimi is on your radar, that's the current number to work with.

GPT-6.1 Terra is now visible on OpenRouter

Four new OpenAI model IDs showed up in our tracker on Oct 27:

openai/gpt-6.1-terra-pro
openai/gpt-6.1-terra-pro:batch
openai/gpt-6.1-terra
openai/gpt-6.1-terra:batch

The batch variants are the standard pattern OpenAI uses for async/discounted batch processing. What's notable here is the "Terra" naming — it's not a suffix OpenAI has used before in models we track. Whether that's a new capability tier, a regional thing, or something else entirely, I genuinely don't know yet. We're just reporting what appeared.

Pricing for these models wasn't in the data we pulled, so I can't give you numbers. Keep an eye on it if you're evaluating GPT-6.1 options — the batch variants in particular tend to come in at a significant discount to the synchronous versions, which matters if your use case can tolerate async turnaround.

What this means practically

A few takeaways if you're making model decisions right now:

DeepSeek Flash users: re-run your cost math with the Oct 31 prices ($0.045 in / $0.50 out). If your workload is output-heavy, you actually got a break. If it's input-heavy, you didn't.
GPT-6.1 Terra: too early to evaluate cost-effectiveness without pricing, but it's worth knowing it exists. The batch variants will likely be the interesting ones once pricing lands.
General: this week is a good illustration of why static pricing docs go stale. If you're building something where API costs matter, you want a live source, not a blog post from three months ago.

We track prices daily across OpenAI, Anthropic, Google, DeepSeek, and xAI. Full current pricing is at llmpricewatch.com.

LLM Pricing Digest: Two New MiniMax Models and Kimi K3.1 Pro Arrive (Week of Oct 20, 2026)

Anders Rasmussen — Sun, 25 Oct 2026 10:02:11 +0000

LLM Pricing Digest: Two New MiniMax Models and Kimi K3.1 Pro Arrive

Weekly pricing intel from LLM Price Watch — tracking Claude, GPT, Gemini, DeepSeek, and Grok.

Quiet week on the price-change front. None of the providers we track moved their rates. No cuts, no increases, nothing to recalculate your cost estimates around.

What did happen: three new models showed up on OpenRouter across two providers. Here's what we spotted and what's worth knowing about each.

MiniMax M3.1 Flash (and a Free Tier)

MiniMax landed two variants this week. The standard minimax/minimax-m3.1-flash appeared on October 23rd, and a free-tier version — minimax/minimax-m3.1-flash:free — had already shown up a few days earlier on October 20th.

The "flash" naming convention signals this is positioned as a fast, lower-cost option rather than a frontier reasoning model. MiniMax hasn't been a household name in Western developer circles, but they've been quietly building capable models and the free tier is a low-friction way to test throughput and output quality before committing budget.

If you're evaluating cheaper inference options for high-volume tasks — summarization, classification, structured extraction — this is worth a benchmark run. The free variant especially gives you no-cost baseline data to work with.

Moonshot AI Kimi K3.1 Pro

The other new arrival is moonshotai/kimi-k3.1-pro, first seen October 24th. Moonshot AI (the company behind Kimi) has been building a reputation in the long-context space. The K3 series appears to be their current-generation lineup, and the "Pro" label suggests this sits above a base or flash variant in their hierarchy.

I don't have pricing details to share yet — it's freshly listed and our tracker is still pulling rate data. Worth checking llmpricewatch.com directly if you want the current numbers before they potentially shift as the model matures on the platform.

What a Quiet Price Week Actually Means

No price changes sounds boring, but it's useful signal. A few things to take from it:

Stability is good for budgeting. If you're mid-project and have cost estimates locked in around any of the five providers we track, nothing changed this week that would break your math.

New models from smaller providers are worth watching. The competitive pressure that's been driving price cuts over the past year often comes from exactly this pattern: a less-established provider lists something capable at a low price point, and the bigger names eventually respond. MiniMax offering a free flash model is part of that longer dynamic.

Free tiers are useful for evaluation, not just savings. The minimax-m3.1-flash:free variant is most valuable as a zero-cost evaluation lane. Run your actual prompts through it, measure latency and quality against whatever you're currently using, then decide if the paid tier makes sense.

What to Watch Next Week

Whether Kimi K3.1 Pro's pricing becomes visible and where it lands relative to comparable mid-tier models
Any response from the major providers to new entrants (price cuts often follow a lag)
Whether MiniMax adds more M3.1 variants — flash models often have a "pro" or "standard" sibling

That's the week. No drama, a few new options to experiment with, and steady prices across the board.

For live pricing across Claude, GPT, Gemini, DeepSeek, and Grok — plus whatever new models surface — the tracker runs daily at llmpricewatch.com.

LLM Pricing Digest: DeepSeek V4, Gemini 3 Flash Prices, and Four New Models This Week

Anders Rasmussen — Sun, 18 Oct 2026 10:02:18 +0000

LLM Pricing Digest: DeepSeek V4, Gemini 3 Flash Prices, and Four New Models This Week

Weekly pricing data from LLM Price Watch — tracking Claude, GPT, Gemini, DeepSeek, and Grok.

This week had a bit of both: a handful of price data points surfacing on newer models, and four new model IDs appearing in our tracker. Here's what actually showed up in the data.

Price Data: What We're Seeing

Gemini 3.8 Flash

Our tracker logged pricing for gemini-3-8-flash at $0.825/M input and $4.125/M output (checked October 18). That output price is notably higher than what Google has trained us to expect from their Flash-tier models. Gemini 2.0 Flash, for comparison, has been sitting at fractions of that on output. If you're running high-output workloads — summaries, long-form generation, agentic loops that produce a lot of tokens — that output cost is worth paying attention to before you swap this in as a drop-in upgrade.

DeepSeek V4 Flash and V4 Pro

Two DeepSeek V4 variants logged pricing on October 12:

deepseek-v4-1-flash: $0.30/M input, $1.20/M output
deepseek-v4-pro: $1.32/M input, $3.96/M output

The Flash variant is extremely cheap. At $0.30 input and $1.20 output, it's in territory that makes it worth testing for anything where cost is the primary constraint — classification, routing, structured extraction, that kind of work. Whether it's good enough for your use case is a different question, but the price is hard to ignore.

The Pro variant at $1.32/$3.96 sits in a more competitive range. It's cheaper than GPT-4o and in the same general neighborhood as Claude Haiku 3.5 on input, though output pricing comparisons depend on your token ratios. If DeepSeek V4 Pro holds up quality-wise, it could be a legitimate contender for mid-tier tasks where you want something better than Flash but don't need to pay frontier prices.

New Models Spotted

Four new model IDs showed up in our tracker this week:

deepseek/deepseek-v4.1-flash-1017 (first seen October 18) — This looks like a dated checkpoint or updated version of the V4 Flash line. DeepSeek has a pattern of releasing timestamped variants, so this is worth watching if you're pinning to specific model versions in production.

qwen/qwen3.7-plus:batch (first seen October 17) — A batch variant of Qwen 3.7 Plus from Alibaba/Qwen. Batch endpoints typically come with meaningful price discounts in exchange for slower turnaround, so if you have async workloads this is worth checking pricing on once it's fully listed.

google/gemini-4-argon and google/gemini-4-argon:batch (both first seen October 14) — This is the most interesting new appearance. Gemini 4 Argon, including a batch variant, showing up in the API. "Argon" as a codename suggests this might be a specific tier or configuration within the Gemini 4 family rather than the flagship. No pricing data in our system yet, so we can't say much about cost positioning.

What This Means Practically

If you're evaluating models right now:

High-volume, cost-sensitive work: DeepSeek V4 Flash at $0.30/$1.20 is worth benchmarking. Just test it against your actual tasks first.
Mid-tier reasoning tasks: DeepSeek V4 Pro enters a crowded field, but the pricing is competitive enough to put it on your shortlist.
Gemini users: The Gemini 3.8 Flash output price is higher than the Flash name might imply — run your token math before assuming it's cheaper than what you're using today.
Keep an eye on Gemini 4 Argon: No pricing yet, but two variants appearing simultaneously (standard + batch) suggests Google is preparing a real release.

Prices change fast. The data above reflects what our tracker detected this week — by the time you read this, some of it may already have shifted.

For live pricing across all five providers, check llmpricewatch.com.

DeepSeek Cuts Prices in Half Mid-Week: This Week in LLM Pricing

Anders Rasmussen — Sun, 11 Oct 2026 10:01:44 +0000

The big story this week is straightforward: DeepSeek halved its prices on two models mid-week. If you're routing any significant volume through either of them, this is worth paying attention to.

The DeepSeek Price Cuts

Our tracker caught two snapshots this week for deepseek-v4-1-flash and deepseek-v4-pro.

deepseek-v4-1-flash

Oct 5: $0.30 input / $1.20 output per million tokens
Oct 10: $0.15 input / $0.60 output per million tokens

deepseek-v4-pro

Oct 5: $1.32 input / $3.96 output per million tokens
Oct 10: $0.66 input / $1.98 output per million tokens

That's a clean 50% reduction across the board for both models, applied between October 5th and October 10th. Not a rounding artifact — both models dropped by exactly half on both input and output.

What does this mean practically? If you've been holding off on deepseek-v4-pro because the cost felt hard to justify against something like GPT-4o or Gemini 1.5 Pro, the math changes now. At $0.66/$1.98, it sits in a much more competitive band for mid-tier inference tasks. And deepseek-v4-1-flash at $0.15 input is genuinely cheap for high-volume, lower-stakes work — summarization, classification, first-pass drafts.

DeepSeek has form here. They've used aggressive pricing as a market share play before, and this follows the same pattern. Whether the prices stay here or drop further is hard to predict, but right now they represent real value if the model quality holds up for your use case.

New Models Spotted

Three new model IDs appeared in our tracker this week.

anthropic/claude-haiku-5.5 and anthropic/claude-haiku-5.5:batch both showed up on October 8th. The batch variant follows Anthropic's established pattern of offering asynchronous batch processing at reduced rates. Haiku has always been Anthropic's speed-and-cost tier, and a 5.5 version suggests incremental improvements over whatever 5.x came before. No pricing anomalies to flag on these yet — we'll track them going forward.

google/gemini-nano-banana-2.1 appeared on October 7th from Google. The "nano" and "banana" naming suggests this is either an experimental model or a specialized lightweight variant — Google has used codename-style identifiers before for models not yet in full public release. Worth watching but nothing concrete to say about it yet beyond the fact that it exists in the API.

The Practical Takeaway

If you have a cost-sensitive workload and you haven't benchmarked DeepSeek V4 Flash recently, now's a reasonable time to do it. At $0.15 per million input tokens, the failure cost of running a quick eval is minimal. For higher-capability needs, V4 Pro at $0.66 input puts it in a range where it's worth comparing directly against your current go-to.

The new Anthropic and Google models are worth keeping an eye on as they mature — particularly Claude Haiku 5.5 if you're already in the Anthropic ecosystem and care about the cost/performance tradeoff at the smaller end.

As always, prices on third-party routers like OpenRouter can lag official announcements or differ slightly from direct API pricing, so verify before committing to anything at scale.

I track these changes weekly over at LLM Price Watch, which checks live pricing across Claude, GPT, Gemini, DeepSeek, and Grok daily. Worth bookmarking if you're making model selection decisions based on cost.

LLM Pricing Digest: Grok-4 Gets a Price Drop, and a Wave of New Models From Qwen, Xiaomi, and Z-AI

Anders Rasmussen — Sun, 04 Oct 2026 10:02:21 +0000

LLM Pricing Digest — Week of September 29, 2026

Weekly data from LLM Price Watch, which tracks live pricing for Claude, GPT, Gemini, DeepSeek, Grok, and more.

The One Real Price Change: Grok-4

This week had one confirmed price movement: grok-4 dropped to $2/million input tokens and $6/million output tokens.

To put that in context, Grok-4 launched as a frontier-tier model, and pricing at that level used to reliably mean $10–$15+ per million on input. At $2 in and $6 out, it's now sitting in territory that makes it genuinely worth comparing against mid-tier options for production workloads — not just for experimentation.

Practically speaking: if you've been routing complex reasoning tasks to GPT-4o or Claude Sonnet and paying ~$5–$15/M input depending on your mix, Grok-4 at $2 input is worth a benchmark run. Output-heavy workloads (long generations, document drafting) will feel the $6/M output price more, but that's still competitive for a model at this capability tier.

No price changes detected for Claude, GPT, or Gemini models this week.

New Models: Qwen, Xiaomi, and Z-AI All Showed Up at Once

Fifteen new models appeared in our tracker on September 30th. They all landed the same day, which usually means a coordinated release push rather than a gradual rollout. Here's the breakdown:

Qwen (5 models)

qwen/qwen3.8-max-prime
qwen/qwen3.8-omni-flash
qwen/qwen3.8-max-0902
qwen/qwen3.8-flash
qwen/qwen3.8-27b
qwen/qwen3.8-27b:free ← free tier variant

Qwen continues to ship variants aggressively. The naming here follows their established pattern: max = higher capability, flash = faster/cheaper, omni = multimodal. The 27B free tier is notable — free access to a 27B model is useful for prototyping or low-volume use cases where cost is the primary constraint.

Xiaomi (3 models)

xiaomi/mimo-v2.6-pro-ultraspeed
xiaomi/mimo-v2.6-flash
xiaomi/mimo-v2.6-pro

Xiaomi's MiMo line is newer to the router ecosystem. Three variants in one drop — ultraspeed, flash, and pro — suggests they're covering the same speed/quality tradeoff spectrum everyone else is. No pricing data to share yet beyond what's in the tracker.

Z-AI / GLM-5.3 (7 models)

z-ai/glm-5.3-prime
z-ai/glm-5.3-flashx
z-ai/glm-5.3-flash
z-ai/glm-5.3-flash:batch
z-ai/glm-5.3
z-ai/glm-5.3:batch

Z-AI's GLM series (from Zhipu AI) is the most prolific of the three this week, with seven variants including batch-mode endpoints. Batch variants matter if you're running offline pipelines — they're typically priced lower in exchange for slower turnaround. Worth checking if you have async workloads that don't need real-time responses.

What This Week Means Practically

If you're actively managing API costs right now:

Re-evaluate Grok-4 if you dismissed it earlier on price. $2 input is a different conversation than where it started.
The Qwen 27B free tier is a legitimate option for dev/test environments or low-stakes internal tools.
GLM-5.3 batch endpoints are worth a look if you're doing bulk processing — batch pricing tends to undercut synchronous endpoints meaningfully.
Xiaomi MiMo is one to watch but probably wait for more community benchmarks before routing production traffic there.

The broader trend continues to hold: the number of capable models available through routing APIs keeps growing, and the middle of the market (the $1–$6/M input range) is getting crowded. That's good for developers with flexibility in model choice.

Full pricing data, historical charts, and alerts are at llmpricewatch.com. The tracker checks OpenRouter daily, so any new price moves on these models will show up there first.

DeepSeek V3 Gets Cheaper Again, Plus Eight GPT-6 Variants and More Hit the API

Anders Rasmussen — Sun, 27 Sep 2026 10:02:06 +0000

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-pro
gpt-6-luna-pro:batch
gpt-6-luna
gpt-6-luna:batch
gpt-6-sol-pro
gpt-6-sol-pro:batch
gpt-6-sol
gpt-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-5.5
claude-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.

LLM Pricing Digest: DeepSeek Quietly Drops a Free Flash Model

Anders Rasmussen — Mon, 21 Sep 2026 08:41:18 +0000

LLM Pricing Digest: DeepSeek Quietly Drops a Free Flash Model

Weekly snapshot from LLM Price Watch — tracking Claude, GPT, Gemini, DeepSeek, and Grok pricing daily.

This was a quiet week on the pricing front. No price cuts, no increases, nothing moved across our five tracked providers. If you're budgeting for an existing integration, your numbers are the same as they were seven days ago.

The one thing worth noting: DeepSeek shipped a new model.

DeepSeek V4 Flash (0731) — Free Tier

We spotted deepseek/deepseek-v4-flash-0731 on September 18th, and it's available on the free tier via OpenRouter. That's the part that actually matters if you're evaluating it for a project.

Here's what I can tell you from the data we have: it's a "flash" variant, which in the current model-naming landscape generally signals a smaller, faster model optimized for lower latency and cost rather than maximum capability. The 0731 date stamp suggests it was trained on or finalized around July 31st of this year. Beyond that, I'm not going to speculate about benchmark numbers or capability claims — we track pricing, not vibes.

What I can say practically: if you've been experimenting with DeepSeek's models and want to test this one without burning through API credits, the free tier is a reasonable way to do that. Run it against your actual use case — summarization, classification, code completion, whatever you're building — and see if the quality holds up before committing.

Why Free-Tier Models Are Worth Watching

Free tiers on OpenRouter tend to come with rate limits and sometimes different routing than paid endpoints, so they're not always a direct proxy for production performance. But they're genuinely useful for:

Prototyping before you decide which model to pay for
Regression testing in CI pipelines where cost adds up fast
Side-by-side evals without having to pre-authorize spend

DeepSeek has been aggressive about making their models accessible, and that's continued here. If you're already using DeepSeek V3 or their R1 series for anything, this flash variant is at least worth a quick benchmark on your workload.

The Bigger Picture This Week

No price changes across Claude, GPT, Gemini, or Grok either. The market has been relatively stable for a few weeks now after a busy stretch of cuts earlier this year. That's not a complaint — stable pricing makes it easier to plan infrastructure costs — but it does mean this digest is shorter than usual.

If you're in the middle of a model selection decision right now, the calculus hasn't shifted this week. The relative cost differences between providers are the same as last week. DeepSeek models remain meaningfully cheaper than OpenAI and Anthropic at the high end, Gemini Flash variants are still the go-to for high-volume, cost-sensitive workloads, and Claude holds its ground for tasks where output quality and instruction-following matter most.

What I'm Watching

A few things I'll be tracking over the next couple of weeks:

Whether DeepSeek V4 Flash gets more detailed specs or a paid pricing tier announced
Any movement on Grok pricing, which has been static for a while
Whether Gemini or OpenAI respond to each other's pricing — historically those two tend to leapfrog

That's it for this week. Short digest, but accurate is more useful than padded.

For live, daily-updated pricing across all five providers, check out llmpricewatch.com. We track input/output token costs and flag changes as soon as they hit.

LLM Pricing Digest: DeepSeek Quietly Drops Two New Flash Models

Anders Rasmussen — Sun, 13 Sep 2026 10:02:23 +0000

LLM Pricing Digest: DeepSeek Quietly Drops Two New Flash Models

Weekly roundup from LLM Price Watch — tracking Claude, GPT, Gemini, DeepSeek, and Grok pricing daily.

This week's data is straightforward: no price changes across any of our five tracked providers. Zero. If you've got a budget locked in for GPT-4o, Claude Sonnet, or Gemini Flash, nothing moved on you.

What did happen is DeepSeek showed up with two new models on OpenRouter, and they're worth knowing about.

What's New from DeepSeek

deepseek/deepseek-v4.1-flash — spotted September 10th.

deepseek/deepseek-v4-flash-vision-exp:batch — spotted September 9th, a day earlier.

Let's talk about what these names actually tell us.

The v4.1-flash model follows DeepSeek's now-familiar naming pattern: a Flash tier is their lower-latency, lower-cost option, comparable in positioning to Gemini Flash or GPT-4o mini. If previous DeepSeek Flash releases are any guide, this is aimed squarely at high-volume, cost-sensitive workloads — think classification, summarization, structured extraction, anything where you're firing off thousands of calls and cost-per-token matters more than raw capability.

The second model is more interesting: deepseek-v4-flash-vision-exp with a :batch suffix. Two things stand out here. First, the vision tag means multimodal input — image understanding on top of text. Second, the exp label means experimental, so treat it accordingly: don't build a production dependency on it yet, but it's worth testing. The :batch variant on OpenRouter typically means asynchronous batch processing at a reduced rate — good for offline pipelines where you don't need a real-time response.

What This Means Practically

If you're currently using DeepSeek for text-only tasks and you've been waiting for a vision-capable option at Flash pricing, this is the moment to run some tests. Vision at Flash cost is a genuinely useful combination — it opens up document parsing, screenshot analysis, and image classification use cases that previously required bumping up to a pricier tier or switching providers entirely.

The batch variant specifically is worth a look if you have any async workloads. Batch endpoints typically offer meaningful discounts over synchronous calls, and if your pipeline can tolerate latency (nightly report generation, bulk document review, etc.), you can often cut costs significantly just by routing to the batch endpoint.

That said, both models are new and one is explicitly experimental. Before committing:

Run your own evals on representative tasks from your actual workload
Check the current pricing on OpenRouter directly — new models sometimes launch at promotional rates that adjust within weeks
Keep an eye on the exp label; experimental models can change, be deprecated, or graduate to stable naming without much notice

The Quiet Week Is Actually Fine

No price changes across Claude, GPT, Gemini, and Grok this week. That's a stable environment for planning. If you've been putting off a cost analysis because prices keep shifting, this is a decent window to benchmark and lock in assumptions.

DeepSeek continues to move fast on model releases. Two models in two days suggests an active release cadence heading into Q4. Whether that means more aggressive pricing pressure on the other providers is something we'll be watching.

I track price changes and new model releases daily across all five major providers. Full live data at llmpricewatch.com.

GPT-6 Astra, Claude Fable, Gemini 3.8: A Busy Week for New LLM Releases

Anders Rasmussen — Tue, 08 Sep 2026 10:37:56 +0000

GPT-6 Astra, Claude Fable, Gemini 3.8: A Busy Week for New LLM Releases

Weekly LLM pricing digest — week of September 5, 2026

No price changes to report this week across the five providers we track (OpenAI, Anthropic, Google, DeepSeek, and xAI). But the model catalog grew noticeably in just a few days, so let's walk through what actually showed up and what it might mean if you're picking a model right now.

OpenAI: GPT-6 Astra lands with a Pro tier and batch variants

The biggest news this week is that OpenAI dropped two new models on September 5th: gpt-6-astra and gpt-6-astra-pro, each with a corresponding :batch variant.

The naming convention here is familiar — the :batch versions are typically lower-cost, asynchronous endpoints suited for workloads where you don't need a real-time response. If you're doing large-scale document processing, evals, or anything that can tolerate a delay, batch is almost always the right call financially.

The split between a base astra and an astra-pro tier suggests OpenAI is continuing the pattern of offering a more capable (and presumably more expensive) variant alongside a leaner everyday option. Worth checking current pricing before assuming which one fits your use case — the gap between base and pro tiers has varied a lot across model generations.

Anthropic: Claude Fable 5.1

Anthropic's new entry, claude-fable-5.1, appeared on September 2nd, again with a batch variant.

The "Fable" name is new in Anthropic's lineup. Whether this sits above, below, or alongside the Sonnet/Opus/Haiku naming structure isn't something our tracker can tell you — we detect models as they appear on OpenRouter, not their internal capability tiers. What we do know is it showed up with batch support from day one, which suggests Anthropic is treating batch as a first-class offering rather than an afterthought.

If you're currently using Claude for any high-volume async workload, it's worth benchmarking fable-5.1 against whatever you're running now.

Google: Gemini 3.8 Flash

gemini-3.8-flash arrived on September 3rd, with batch support included.

The Flash line has consistently been Google's price-performance workhorse — fast, cheap, and good enough for a wide range of tasks. 3.8 continuing that series is a reasonable assumption, though again, actual pricing on our tracker is what should drive your decision, not the name alone.

If you're already using Gemini 2.x Flash for something latency-sensitive or cost-sensitive, this is probably worth a quick test.

xAI: Grok 4.3 Batch

xAI added grok-4.3:batch on September 4th. This appears to be a batch variant of an existing model rather than a completely new release, which fits the pattern of providers progressively rolling out async options for their model lineup.

What this week means practically

Five new model IDs in four days across four different providers, with zero price changes on existing models. A few things stand out:

Batch variants are becoming standard. Every new model this week came with a batch option. If you're not already routing eligible workloads to batch endpoints, you're likely leaving money on the table.
The catalog is expanding faster than pricing is changing. That means the cost landscape for existing models is stable right now, but the choice of which model to use is getting more complex. It's worth revisiting your model selection if you haven't done so recently.
None of these models have a pricing history yet. First-week pricing on a new model sometimes gets revised. It's not unusual to see adjustments in the first few weeks after launch.

We'll be tracking all of these as they settle in. If prices shift on any of the new models — or on anything else across our five providers — it'll show up in next week's digest.

For live pricing on all of these models and more, check LLM Price Watch.

If you want the real story of how this tracker (and its sister site, StackIndex AI) got built — domain traps, design mistakes, the works — it's now a course: StackIndex Academy. First 3 chapters are free, no signup.

GPT-6 Astra, Claude Fable, Gemini 3.8: A Busy Week for New LLM Releases (No Price Changes Though)

Anders Rasmussen — Sun, 06 Sep 2026 10:01:47 +0000

GPT-6 Astra, Claude Fable, Gemini 3.8: A Busy Week for New LLM Releases

Weekly digest from LLM Price Watch — tracking live pricing across OpenAI, Anthropic, Google, DeepSeek, and xAI.

No price changes this week across any of the five providers we track. Zero. But the model catalog? That got noticeably busier. Between September 2nd and 5th, we spotted nine new model IDs across four providers. Here's what showed up and what's worth paying attention to.

OpenAI: GPT-6 Astra (and a Pro tier)

The biggest names to land on our radar are openai/gpt-6-astra and openai/gpt-6-astra-pro, both spotted on September 5th. Both also have batch variants (gpt-6-astra:batch and gpt-6-astra-pro:batch).

The GPT-6 naming is a notable jump — we've been sitting with GPT-4 variants for a long time in practical use. The "Astra" label has been associated with Google's multimodal work in the past, so it's an interesting name choice from OpenAI. The Pro tier suggests a familiar tiered pricing structure is coming, similar to what we've seen with other model families.

Practically speaking: if you're building something today, I'd hold off on committing to these until pricing is confirmed and benchmarks are out. Batch variants being available at launch is a good sign — batch processing has been one of the more reliable ways to cut costs on longer, async workloads.

xAI: Grok 4.3 Batch

x-ai/grok-4.3:batch appeared on September 4th. This is a batch-only variant, no standard inference version detected yet. xAI has been iterating fairly quickly on the Grok 4 line, and adding batch support makes sense if they're targeting workloads where latency isn't critical — think evals, document processing, data pipelines.

Google: Gemini 3.8 Flash

google/gemini-3.8-flash and its batch counterpart showed up on September 3rd. The Flash line from Google has generally been their cost-efficient, faster-response tier — good for high-volume use cases where you don't need the full Gemini Pro/Ultra capability. Gemini 3.8 continuing that Flash tradition makes sense. If past Flash pricing holds, this could be one of the cheaper options in the new generation lineup once pricing is confirmed.

Anthropic: Claude Fable 5.1

anthropic/claude-fable-5.1 and anthropic/claude-fable-5.1:batch appeared on September 2nd, making it the earliest arrival in this week's batch.

The "Fable" name is new for Anthropic — we haven't tracked a Fable line before. It's not clear yet where this sits relative to Sonnet, Haiku, or Opus in terms of capability or price point. The name might suggest a model tuned for narrative or creative tasks, but that's speculation. Worth watching for the official positioning.

What This Week Actually Means

With no price changes, your current cost modeling stays the same. Nothing got cheaper or more expensive in the existing lineup.

The main thing to flag is that four providers dropped new model IDs in a four-day window. That's a lot of surface area appearing at once. A few practical notes:

Don't route production traffic to these yet unless you've tested them. New model IDs appearing in the API doesn't always mean they're fully stable or that pricing is final.
Batch variants at launch is genuinely useful — it signals the providers are thinking about cost-sensitive workloads from day one.
The GPT-6 naming is the most significant signal here. If OpenAI is moving to a 6.x generation label, expect Anthropic and Google to respond with their own positioning in the coming weeks.

I'll be watching for actual pricing data on all of these. Once numbers are confirmed, I'll do a proper comparison.

For live pricing across all tracked models and providers, check llmpricewatch.com.

GPT-6 Astra, Claude Fable, Gemini 3.8: A Busy Week for New LLM Releases

Anders Rasmussen — Sat, 05 Sep 2026 19:04:05 +0000

GPT-6 Astra, Claude Fable, Gemini 3.8: A Busy Week for New LLM Releases

Weekly digest from LLM Price Watch — tracking real pricing data across OpenAI, Anthropic, Google, DeepSeek, and xAI.

No price changes to report this week across our tracked providers. But the model catalog grew significantly — nine new model IDs showed up in our tracker between September 2nd and 5th, 2026, spanning all four of the big players we watch. Here's what appeared and what's worth noting practically.

OpenAI: GPT-6 Astra (and a Pro tier)

The biggest news by name alone: OpenAI dropped two new model families on September 5th — gpt-6-astra and gpt-6-astra-pro, each with a corresponding :batch variant.

The batch variants are the immediate practical signal here. OpenAI has consistently priced batch inference at a meaningful discount (typically around 50% off standard rates), so if you're running non-latency-sensitive workloads — evaluations, document processing, bulk classification — the batch endpoints are worth watching closely once pricing is confirmed.

The split between a base gpt-6-astra and gpt-6-astra-pro suggests a tiered capability structure, similar to how the GPT-4o / GPT-4o-mini split played out. Whether the Pro variant is meaningfully better for reasoning-heavy tasks or just a marketing distinction remains to be seen. I'd hold off on routing production traffic until pricing and benchmark data are clearer.

xAI: Grok 4.3 Batch

On September 4th, grok-4.3:batch appeared — notably, just the batch variant, with no corresponding standard model showing up in our tracker this week. That's an interesting release pattern. It could mean the base grok-4.3 was already present in the catalog, or that xAI is leading with batch access before a broader rollout. Either way, if you're already using Grok models for async workloads, this is worth checking against your current setup.

Google: Gemini 3.8 Flash

September 3rd brought gemini-3.8-flash and its batch counterpart. The Flash line from Google has generally been their price-performance sweet spot — faster and cheaper than Pro variants, with quality that's often sufficient for structured output tasks, summarization, and classification.

If the 3.8 Flash follows the same pricing trajectory as earlier Flash models, it could become a solid default for high-volume, cost-sensitive applications. The batch variant showing up simultaneously with the standard model is a good sign — it suggests Google is treating batch access as a first-class feature rather than an afterthought.

Anthropic: Claude Fable 5.1

The most intriguing naming this week: claude-fable-5.1 appeared on September 2nd, again with a batch variant. "Fable" isn't a naming convention we've seen from Anthropic before — their previous lines have been Haiku, Sonnet, and Opus. A new name likely means a new positioning, possibly a specialized or fine-tuned variant rather than a general-purpose tier.

Without more documentation from Anthropic, it's hard to know where Fable sits in their lineup — whether it's meant to replace something, complement the existing tiers, or target a specific use case. The 5.1 version number suggests it's not a ground-up new model but an iteration of something in the Claude 5 family.

What This Means Practically

A week with nine new model IDs and zero price changes is a week for watching, not necessarily acting. A few things I'd suggest:

Don't rush to migrate to any of these until pricing is confirmed and published. New model IDs on OpenRouter don't always come with finalized pricing on day one.
Batch variants are worth bookmarking across all four providers. If your workload tolerates async processing, batch endpoints typically offer the best cost efficiency available.
The Anthropic "Fable" naming is unusual enough to keep an eye on — it may signal a meaningful capability shift or specialization that affects how you'd evaluate it against Sonnet or Haiku.

I'll be tracking pricing as it settles across all nine of these models in next week's digest.

For live pricing and model tracking across Claude, GPT, Gemini, DeepSeek, and Grok, visit llmpricewatch.com.

New Batch Models From DeepSeek, Google, and OpenAI: LLM Pricing Digest

Anders Rasmussen — Sun, 30 Aug 2026 10:01:38 +0000

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:batch
deepseek/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:batch
openai/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.