LLM Market Snapshot for Coding Agents (September 2026)

Last updated: Oct 1, 2026

Reading time: ≈20 minutes

Part of: the AI FinOps section.

Purpose: Answer two questions with dated facts: are LLM costs rising or falling, and what does agent work cost on each provider today. This page is a snapshot, not a live price list.

Data snapshot date

Every price and quota on this page was read at the vendor’s own documentation on 2026-09-30, unless a row says otherwise. Vendors reprice without notice and rename models between releases. Re-verify at the linked source before any purchase decision.

How the numbers were checked. The starting point was an AI-generated market summary. Each claim was then re-read at a primary source: vendor pricing pages, help centers, API documentation, and changelogs. The reading tool returns a summarized version of each page, so an exact figure can still carry a transcription error. openai.com and help.openai.com refused automated reads during verification; OpenAI figures come from developers.openai.com and OpenAI’s learn.chatgpt.com pricing page instead. Where only a secondary source was available, the row says so.

1. The trend: quotas shrink while token prices fall

Three kinds of change get reported as “prices going up”. Keep them apart:

API price: the price per million tokens changes.
Quota: a subscription includes less (or more) usage at the same price.
Conditions: windows, model weighting, retired models, or which clients may use a plan.

Date	Provider	Change	Type	Source
2025-05-08	Google	Implicit caching on Gemini 2.5: 75% discount on cached input	API, down	Google Developers Blog
2025-06-17	Google	Gemini 2.5 Flash GA: input $0.15 → $0.30, output $3.50 → $2.50 per 1M tokens	API, mixed	Google Developers Blog
2025-07-07	Google	Gemini Batch API at 50% of the standard rate	API, down	Google Developers Blog
2025-08-21	DeepSeek	New V3.1 pricing; off-peak discounts end on 2025-09-05, 16:00 UTC	API, off-peak discount removed	DeepSeek news
2025-12-09	Mistral	Devstral 2 launched free on the API, with $0.40 / $2.00 announced for later	API	Mistral
2026-01-27	Mistral	Devstral 2 becomes paid ($0.40 / $2.00); free access limited to the Experiment plan	API, up	Mistral
2026-05-15	xAI	Several Grok models retired; reasoning traffic redirected to `grok-4.3` at $1.25 / $2.50	Conditions	xAI docs
2026-06-01	Cursor	Teams Standard $40 per seat monthly ($32 annual), Premium $120 ($96 annual); renewals from 2026-07-01	Subscription price	Cursor
2026-06	Anthropic	The help article, updated in mid-June, pauses the separate monthly credit announced for Agent SDK, `claude -p`, and third-party apps. Read on 2026-09-30: that usage still draws from the subscription’s limits	Conditions	Claude Help Center
2026-07-30	Z.AI	GLM Coding Plan moves to credits weighted by model and time of day	Quota	Z.AI notice
2026-08-13	DeepSeek	V4-Pro GA with peak and off-peak pricing, off-peak at half the peak rate, effective 2026-08-16	API, down off-peak	DeepSeek updates
2026-09-01	Fireworks	Newly launched US-only models priced at 1.5x the base serverless price	API, regional premium	Fireworks docs
2026-09-22	Anthropic	Claude Opus 5.5: cache read $0.20 per 1M tokens, 60% lower than its predecessor ($0.50)	API, down	Anthropic
2026-09-30 (read)	Anthropic	Claude Fable 5.1 listed at $10 / $50 with cache reads at $0.25, 0.025x the input price. Anthropic recommends Opus 5.5 for most workloads and Fable 5.1 for demanding reasoning and long-horizon agentic work	API	Anthropic models overview, Anthropic pricing
2026-09-30 (read)	OpenAI	GPT-6 Sol and GPT-6.1 Sol listed at $2 / $10, one fifth of GPT-6 Astra’s $10 / $50. The launch date and the previous price were not readable at OpenAI	API	OpenAI model page
2026-09-29	OpenAI	New Pro $500 plan, the only Pro plan with Astra Ultrafast. Pro $200: included usage in ChatGPT Work and Codex falls from 20x to 10x the Plus allowance on 2026-10-30, GPT-6 Pro chat messages from 200 to 100 per week, price unchanged	Quota, down	Tiers: OpenAI pricing. Pro $200 change: The Next Web, 2026-09-29 (secondary; OpenAI’s help pages were not readable)

Anthropic’s pricing documentation also states that a planned increase of Claude Sonnet 5 to $3 / $15, scheduled for 2026-09-01, will not occur. Sonnet 5 stays at $2 / $10.

Reading of the trend. Most documented API changes lower the bill: cheaper cache reads (Gemini 2.5 implicit caching, Opus 5.5), batch and off-peak discounts, and a cancelled increase (Sonnet 5). The increases are targeted: Gemini 2.5 Flash input, the end of DeepSeek’s off-peak discounts in 2025, the Devstral 2 launch period, and regional premiums. On the subscription side, the dated changes are Z.AI’s switch to weighted credits in July 2026 and OpenAI’s halving of the Pro $200 allowance at the same price. Across the changes listed here, the reductions hit subscription quotas and conditions; per-token list prices rose only in the targeted cases above.

2. API prices

USD per 1 million tokens, standard tier, read on 2026-09-30. “n/a” means no cached-input price was found, not that caching is free.

Provider and model	Input	Cached input	Output	Modifiers	Source
Anthropic Claude Fable 5.1	$10.00	$0.25	$50.00	Batch -50%; cache read 0.025x input; 5-minute cache write $12.50	Anthropic pricing
Anthropic Claude Opus 5.5	$4.00	$0.20	$20.00	Batch -50%; US-only inference 1.1x; fast mode $8 / $40	Anthropic pricing
Anthropic Claude Sonnet 5.5	$2.00	$0.20	$10.00	Batch -50%; 5-minute cache write $2.50	Anthropic pricing
OpenAI GPT-6 Astra	$10.00	$1.00	$50.00	Batch and Flex -50%; Fast 2x; cache write $12.50	OpenAI model page
OpenAI GPT-6 Sol and GPT-6.1 Sol	$2.00	$0.20	$10.00	Cache write 1.25x input; above 272K input tokens, input and cache 2x, output 1.5x	OpenAI model page, OpenAI pricing
OpenAI GPT-6 Luna	$0.10	$0.01	$0.50	Same modifiers as Sol	OpenAI model page
Google Gemini 3.1 Flash-Lite	$0.25	n/a	$1.50	Batch $0.125 / $0.75; preview at launch	Gemini API pricing
DeepSeek `deepseek-v4-pro`, peak	$1.32	$0.044	$3.96	Off-peak half price: $0.66 / $0.022 / $1.98	DeepSeek pricing
DeepSeek `deepseek-flash`, peak	$0.30	$0.006	$1.20	Off-peak half price: $0.15 / $0.003 / $0.60	DeepSeek pricing
Mistral Devstral 2	$0.40	n/a	$2.00	Devstral Small 2: $0.10 / $0.30	Mistral
xAI Grok 4.7	$2.00	n/a	$6.00	US regional endpoint +10%	xAI docs
Moonshot Kimi K3	$3.00	$0.30	$15.00	Cache write $3.00 (5-minute TTL) or $6.00 (1-hour TTL)	Kimi platform pricing
Z.AI GLM-5.3	$1.40	$0.26	$4.40		Z.AI pricing
Together Qwen3.6-Plus	$0.50	n/a	$3.00	Qwen3.7-Plus listed at $0.32 / $1.28	Together pricing
Fireworks DeepSeek V4.1 Flash	$0.30	$0.006	$1.20	Batch -50%; Priority +25%; launched US-only models 1.5x	Fireworks pricing
Groq GPT-OSS 120B	$0.15	$0.075	$0.60		Groq model page
OVHcloud AI Endpoints GPT-OSS 120B	€0.08	n/a	€0.40	EUR, as published	OVHcloud catalog
OpenRouter (gateway)	Provider list price	Per endpoint	Provider list price	No inference markup; credit purchase fee 5.5% (Standard) or 8% (Business); bring-your-own-key free up to $25,000 of list-price inference per month ($200,000 on Enterprise), then 5%	OpenRouter pricing

Per-token prices do not compare across tokenizers. Anthropic documents that Claude 4.7 and later models produce roughly 30% more tokens for the same text than earlier Claude models. Tokenizers also differ between vendors, so the same prompt can produce a different token count on each, and a cheaper per-token price can still produce a larger bill.

3. Subscription quotas

No vendor below publishes a subscription allowance as a token budget, except where the row says so. A multiplier (“5x Pro”) compares one plan with another, and model, effort, and context size weight the consumption.

Offer	Price	Published allowance	Conditions	Source
Claude Pro	$20 monthly, $17 annual	At least 5x the Free plan per 5-hour session	5-hour sessions plus weekly limits; Claude and Claude Code share the same limits; usage credits at API rates after the limit	Claude pricing, Claude Code with Pro or Max
Claude Max 5x / 20x	$100 / $200 monthly	5x / 20x the Pro per-session allowance	Weekly and monthly caps possible	Max plan
Claude Team Standard / Premium	$25 / $125 per seat monthly, $20 / $100 annual; minimum 2 members	1.25x / 6.25x the Pro per-session allowance	Individual limits per seat	Team plan
Claude Enterprise, usage-based	$20 per seat per month, billed annually	Usage billed at API rates	See subscription strategy	Claude pricing
ChatGPT Go / Plus	$8 / $20 monthly	Plus is the base unit for Pro multipliers	Plus has a 5-hour limit in Work and Codex	OpenAI pricing
ChatGPT Business	$20 per user annual, $25 monthly	Not published in tokens	Standard Business has a 5-hour limit	OpenAI pricing
ChatGPT Pro $100 / $200 / $500	$100 / $200 / $500 monthly	Pro $200: 10x Plus from 2026-10-30 (20x before). Pro $100 and Pro $500 multiples were not readable on an OpenAI page	Pro plans have no 5-hour limit; among Pro plans, Astra Ultrafast is only on Pro $500	OpenAI pricing, The Next Web
Cursor Pro / Pro+ / Ultra	$20 / $60 / $200 monthly	Not published in tokens		Cursor pricing
GitHub Copilot Pro / Pro+ / Max	$10 / $39 / $100 monthly	1,500 / 7,000 / 20,000 AI credits per month; 1 credit = $0.01	Each allowance splits into base and flex credits (1,000 + 500 on Pro)	GitHub Copilot plans
Mistral Le Chat Pro / Team	$14.99 monthly / $24.99 per user ($50 minimum)	Pro includes $15 per month of API credits	Vibe usage subject to fair-use limits, no numbers published	Mistral pricing
Z.AI GLM Coding Plan	From $18 monthly	Lite / Pro / Max: 10,000 / 60,000 / 140,000 credits per week; 2,000 / 12,000 / 28,000 per 5 hours	Peak Monday to Friday 14:00 to 18:00 SGT (UTC+8); GLM-5.3 costs 3x credits at peak vs 1x off-peak; token examples assume a 95% cache hit rate	Z.AI Coding Plan
Alibaba Model Studio Coding Plan	¥200 monthly	6,000 requests per 5 hours, 45,000 per week, 90,000 per month	Quota counted per model call, not per prompt	Alibaba Cloud
OpenCode Go	$10 monthly	See the guide’s OpenCode Go section	Single seat; a Go Plus tier at $40 is also listed	OpenCode Go
Cerebras Code Pro / Max	$50 / $200 monthly	Up to 24M / 120M tokens per day, the only allowance here published in tokens	Both tiers shown as sold out on 2026-09-30; whether cached tokens count toward the daily figure is not stated	Cerebras Code

Third-party clients on a Claude subscription. Anthropic’s legal and compliance page states that third-party developers may not route requests through Free, Pro, or Max plan credentials on behalf of their users, and that a subscriber may sign in to Claude Code with their own plan (Claude Code legal and compliance). Press coverage dates specific enforcement actions to January and April 2026; no Anthropic page with those dates was found.

4. Capability: read the benchmark conditions first

A benchmark score measures a model, a harness, an effort setting, and a number of attempts together. Two scores are comparable only when all four match and the benchmark version is the same.

Model	Benchmark	Score	Conditions stated	Measured by	Date
Claude Opus 5.5	Terminal-Bench 4.0	66.4% (±2.6)	xhigh effort, Claude Code harness	Anthropic	2026-09-22
Claude Sonnet 5.5	Terminal-Bench 4.0	70.6%	Effort and harness not stated in the text read	Anthropic	2026-09-28
Gemini 3.5 Flash	Terminal-Bench 2.1	76.2%	Different benchmark version: not comparable with the rows above	Google	2026-05-19

Sources: Claude Opus 5.5, Claude Sonnet 5.5, Gemini 3.5.

All three are vendor-reported. Do not read the first two rows as “Sonnet beats Opus” until both conditions are published. No independent benchmark run with one harness across all providers was found. For a buying decision, run your own tasks through the same harness on each candidate, as described in agent evaluation.

5. Cost of one reference workload

The workload is a hypothesis chosen to make providers comparable, not a measured average: one day of agent work with 20M input tokens, 80% of them served from cache (16M cached, 4M uncached), and 0.4M output tokens.

Cost = 4 × input price + 16 × cached price + 0.4 × output price. When no cached price was found, all 20M input tokens are billed at the full rate, which makes that row an upper bound. Cache writes are excluded because they depend on how prompts are built.

Model	Cost for the day
OpenAI GPT-6 Astra	$76.00
Anthropic Claude Fable 5.1	$64.00
xAI Grok 4.7 (no cached price found, upper bound)	$42.40
Anthropic Claude Opus 5.5	$27.20
Moonshot Kimi K3	$22.80
OpenAI GPT-6 Sol and GPT-6.1 Sol	$15.20
Anthropic Claude Sonnet 5.5	$15.20
Z.AI GLM-5.3	$11.52
Mistral Devstral 2 (no cached price found, upper bound)	$8.80
DeepSeek `deepseek-v4-pro`, peak / off-peak	$7.57 / $3.78
Groq GPT-OSS 120B	$2.04
DeepSeek `deepseek-flash`, peak / off-peak	$1.78 / $0.89
OVHcloud GPT-OSS 120B (EUR, no cached price found)	€1.76
OpenAI GPT-6 Luna	$0.76

One hypothetical day of agent work, 20M input tokens with 16M served from cache and 0.4M output tokens, priced at list rates read on 2026-09-30, ranging from $76.00 on OpenAI GPT-6 Astra to $0.76 on OpenAI GPT-6 Luna for the same token count.

The chart leaves out the three upper-bound rows (no cached price found) and Claude Fable 5.1, which was added to the table after the chart was generated.

This table compares bills for the same token count. It does not compare work done. A model that fails, loops on tool calls, or needs three attempts costs more per accepted task than its row suggests, and the tokenizer difference noted in §2 changes the token count itself. Measure cost per accepted task on your own workload with the method in AI unit economics.

6. Where the data goes

Provider	What the documentation states
Anthropic API	Inference geography: only `us` and `global`; workspace geography: only `us` (data residency). Inputs and outputs from commercial products are not used for training by default (privacy center)
OpenAI API	European endpoint `eu.api.openai.com`, which requires approval and a Modified Retention amendment, with a 10% regional processing premium where available. System data (account, metadata, usage) may be processed outside the region. API data is not used for training unless you opt in (your data)
Google Gemini API	Paid-tier content is not used to improve Google’s products; free-tier content may be (Gemini API pricing). The Vertex AI residency page lists France and Germany regions, but which commitment and which models they cover was not established (Vertex data residency)
DeepSeek API	Personal data is collected, processed, and stored in the People’s Republic of China, with a GDPR Article 27 representative for the EU. No EU residency option was found (privacy policy)
OpenRouter	In-region EU routing on Business and Enterprise plans; zero-data-retention routing on every plan (OpenRouter pricing)
xAI, Fireworks	Regional premiums: xAI +10% on the US endpoint, Fireworks 1.5x on newly launched US-only models

Using an LLM API under the GDPR involves a data processing agreement, a legal basis, and control of subprocessors and international transfers. Choosing an EU region, where a provider offers one, narrows the transfer question but does not replace the rest.

7. Claims that did not survive verification

List prices in the AI-generated summary held up almost everywhere they could be checked. Its errors clustered in dates, plan conditions, and model attribution:

Claim in the summary	What the primary source says
Since June 15, 2026, Agent SDK and `claude -p` usage has its own credit	Announced, then paused. That usage still draws from subscription limits
Cerebras Code gives 24M tokens a day for $50	The figure is real, but the source post dates from November 2025 and both paid tiers were sold out on 2026-09-30
DeepSeek off-peak means nights and weekends at half price	Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays; everything else, weekends included, is off-peak
OpenRouter zero data retention is a Business and Enterprise feature	Zero-data-retention routing is on every plan; EU in-region routing is the Business and Enterprise feature
Gemini 3.5 scores 76.2% on Terminal-Bench 2.1	The score belongs to Gemini 3.5 Flash
Scaleway Generative APIs start at €0.15 input, L4 GPU at €0.93 per hour	Not found on Scaleway’s price pages, which showed different figures on 2026-09-30

How the price snapshot was checked: claims from an AI-generated summary were re-read at vendor pricing pages, help centers, API documentation and changelogs; list prices held almost everywhere, while errors clustered in dates, plan conditions and model attribution.

8. Refresh rule

Date every figure. A price without the date it was read is not usable.
Re-verify at the primary source before quoting a figure in a purchase decision, a budget, or a published comparison.
Refresh this page each quarter, and immediately after a vendor event that changes plans (a developer conference, a model launch, a plan email).
Keep dated events out of the table unless a primary or named secondary source carries the date. An AI-generated summary does not count as a source.

What this snapshot could not establish: Pro $100 and Pro $500 multiples on an OpenAI page; GPT-6 Sol’s launch date and previous price; prices of Gemini models released after 3.1 Flash-Lite; DeepSeek’s February 2025 off-peak discount at a DeepSeek page; per-model Azure and AWS Bedrock prices; Groq batch rules; Cerebras shared-inference prices; Scaleway token and GPU prices; the regions and models covered by Vertex AI residency; volume discounts, which are generally negotiated privately.