What does $1,000 a month of LLM API buy in 2026? Three real workloads, priced out
A support chatbot, a coding agent and a document-extraction batch: explicit token assumptions and formulas show how many units $1,000 buys on eight models, and how caching, output volume, reasoning tokens and price flapping can double the bill.
1. What $1,000 a month means in token terms
Every pricing page quotes dollars per million tokens, and almost nobody budgets in that unit. Budgets are written in the unit of the business: conversations handled, tasks completed, documents processed. This article converts our price data (as of 2026-10-05) into those units. We define three workloads with explicit token counts, price each one on eight models across three tiers, and ask a single question: how many units does $1,000 buy, and what quality do you give up to get more of them?
Every number below comes from one formula, applied to the current rows of our pricing database:
cost = (T_fresh × P_in + T_cached × P_cache + T_out × P_out) / 1,000,000
units per $1,000 = 1,000 / cost
Prices are USD per 1M tokens and mostly reflect the default endpoint at OpenRouter. Quality is shown as the Artificial Analysis Intelligence Index (AA) and the AA Coding Index, using the latest value stored for each model (the 2026-09-08 snapshot; AA has re-based the index several times this year, so these scores are not comparable with older figures). Both are general benchmarks. They say nothing about your specific task, so read the quality columns as a prior, not a verdict.
2. Three workloads, stated explicitly
The assumptions are in the table so you can swap in your own. T_fresh is input billed at the normal input price, T_cached is input served from prompt caching, T_out is output.
| Workload | Unit | Fresh input | Cached input | Output |
|---|---|---|---|---|
| A. Support chatbot | 1 conversation (6 turns) | 15,000 | 0 | 1,500 |
| B. Coding agent | 1 task (30 model calls) | 180,000 | 1,620,000 | 45,000 |
| C. Document batch | 1 document (~45 pages) | 30,000 | 0 | 800 |
A, support chatbot. A 1,500-token system prompt with retrieved knowledge-base snippets, then six turns of about 100 user tokens and 250 reply tokens. Each turn re-sends everything before it, so the six inputs sum to roughly 14,850 tokens; we round to 15,000. Output is 6 × 250 = 1,500. The baseline uses no caching, which is conservative, and section 7 adds it back.
B, coding agent. One task is 30 model calls with an average context of 60K tokens, so 1.8M input tokens per task. Agent loops re-send a nearly identical prefix, so we assume 90% is served from cache (1.62M cached, 180K fresh). Each call emits about 1,500 tokens, 45,000 in total. Real tasks vary by an order of magnitude in both directions; this is a mid-sized one.
C, document batch. A report of about 45 pages is roughly 30,000 tokens. The model returns an 800-token structured summary or JSON extraction. One call per document, no caching.
Here is the formula worked through for Claude Opus 5 ($5.00 input, $0.50 cached, $25.00 output). Workload A: (15,000 × 5 + 0 × 0.5 + 1,500 × 25) / 1,000,000 = 0.075 + 0.0375 = $0.1125 per conversation, so 1,000 / 0.1125 = 8,889 conversations. Workload B: (180,000 × 5 + 1,620,000 × 0.5 + 45,000 × 25) / 1,000,000 = 0.90 + 0.81 + 1.125 = $2.835 per task, or 353 tasks. To adapt it, change the three token counts in the table to what your logs show and rerun the same arithmetic for your candidate model.
3. Eight models across three tiers
Frontier is Claude Opus 5 and GPT-5.6 Sol, the third and fifth highest AA scores among every model in our database that has both a price and an AA score. We skip the other three at the top (Claude Fable 5.1, GPT-6 Astra and Claude Fable 5), which are all priced at $10/$50 and would simply double Opus 5's bill. Mid is Claude Sonnet 5, Gemini 3.8 Flash and Qwen3.8 Max. Budget is GLM 5.3 Flash, DeepSeek V4 Pro and MiniMax M2.7.
| Model | Tier | Input $/M | Cached $/M | Output $/M | AA | Coding | Context | 30-day input range |
|---|---|---|---|---|---|---|---|---|
| Claude Opus 5 | Frontier | $5.00 | $0.50 | $25.00 | 50.7 | 78.0 | 1M | $5.00 |
| GPT-5.6 Sol | Frontier | $2.00 | $0.20 | $10.00 | 47.1 | 77.4 | 1.05M | $2.00 |
| Claude Sonnet 5 | Mid | $2.00 | $0.20 | $10.00 | 38.4 | 71.5 | 1M | $2.00 |
| Gemini 3.8 Flash | Mid | $0.75 | $0.075 | $3.75 | 41.2 | 76.3 | 1M | $0.75 |
| Qwen3.8 Max | Mid | $2.00 | $0.25 | $6.00 | 40.3 | 71.8 | 1M | $2.00 |
| GLM 5.3 Flash | Budget / open-weight | $0.15 | $0.03 | $0.50 | 41.9 | 71.5 | 1M | $0.04–0.15 |
| DeepSeek V4 Pro | Budget / open-weight | $0.2088 | $0.0174 | $0.4176 | 30.9 | 59.4 | 1M | $0.209–1.60 |
| MiniMax M2.7 | Budget / open-weight | $0.21 | $0.042 | $0.84 | 23.2 | 52.6 | 205K | $0.21–0.30 |
We checked that input and output prices are in a normal relationship for every row: output is 5× input for Opus 5, Sol, Sonnet 5 and Gemini, 3× for Qwen, about 3.3× for GLM 5.3 Flash, 4× for MiniMax and 2× for DeepSeek. We left out DeepSeek V4 Flash for exactly this reason: its current record shows $0.03 input and $1.28 output, a 43× ratio far outside the others, which looks like an endpoint artefact rather than a real list price.
The last column is the more important warning. Prices for the first four models did not move in 30 days. Qwen3.8 Max has not appeared in our daily snapshots since 2026-09-04; OpenRouter now lists it as Qwen3.8 Max 0902, at the same $2/$6, which has not moved since it appeared on 09-05. The three budget models did: GLM 5.3 Flash input ranged from $0.04 to $0.15 (today is the top of that range), DeepSeek V4 Pro from $0.2088 to $1.60 (today is the bottom), and MiniMax M2.7 from $0.21 to $0.30 (today is the bottom). Those swings are the router's default endpoint changing, not list-price cuts. For these three, read the results as a range, which section 7 quantifies. Prices do also move for real: GPT-5.6 Sol went from $5/$30 on 2026-07-09 to $2/$10 on 2026-08-22, a 60% cut on input and 67% on output.
4. Workload A: the support chatbot
| Model | AA | Cost / conversation | Conversations / $1,000 |
|---|---|---|---|
| Claude Opus 5 | 50.7 | $0.1125 | 8,889 |
| GPT-5.6 Sol | 47.1 | $0.0450 | 22,222 |
| Claude Sonnet 5 | 38.4 | $0.0450 | 22,222 |
| Gemini 3.8 Flash | 41.2 | $0.0169 | 59,259 |
| Qwen3.8 Max | 40.3 | $0.0390 | 25,641 |
| GLM 5.3 Flash | 41.9 | $0.00300 | 333,333 |
| DeepSeek V4 Pro | 30.9 | $0.00376 | 266,071 |
| MiniMax M2.7 | 23.2 | $0.00441 | 226,757 |
The spread is enormous. Opus 5 costs $0.1125 per conversation, so $1,000 covers 8,889 of them. GLM 5.3 Flash covers 333,333, which is 37.5× more conversations for a model with a 17% lower AA (41.9 vs 50.7). Whether that gap matters depends on how hard your conversations are. If most are order-status questions answered from retrieved text, you are paying a premium for headroom you do not use.
Two comparisons stand out. Sonnet 5 and GPT-5.6 Sol have identical prices ($2/$10) and therefore identical cost, 22,222 conversations, but Sol has the higher AA, 47.1 against 38.4. On these two indexes Sonnet 5 buys you nothing extra at the same price, so choose it only for reasons the indexes do not capture. And Gemini 3.8 Flash, at 59,259 conversations and AA 41.2, is beaten on both axes by GLM 5.3 Flash (333,333 and 41.9) at today's prices. That conclusion depends on GLM's current price holding, since it sits at its 30-day high rather than its low.
5. Workload B: the coding agent
| Model | Coding | Cost / task | Tasks / $1,000 | Cost / task, no cache | Tasks, no cache |
|---|---|---|---|---|---|
| Claude Opus 5 | 78.0 | $2.835 | 353 | $10.125 | 99 |
| GPT-5.6 Sol | 77.4 | $1.134 | 882 | $4.050 | 247 |
| Claude Sonnet 5 | 71.5 | $1.134 | 882 | $4.050 | 247 |
| Gemini 3.8 Flash | 76.3 | $0.4253 | 2,352 | $1.519 | 658 |
| Qwen3.8 Max | 71.8 | $1.035 | 966 | $3.870 | 258 |
| GLM 5.3 Flash | 71.5 | $0.0981 | 10,194 | $0.2925 | 3,419 |
| DeepSeek V4 Pro | 59.4 | $0.0846 | 11,825 | $0.3946 | 2,534 |
| MiniMax M2.7 | 52.6 | $0.1436 | 6,962 | $0.4158 | 2,405 |
Caching is the story here. Without it, Opus 5 costs $10.125 per task and $1,000 buys 99 tasks; with 90% cache hits it costs $2.835 and buys 353. For the frontier models caching removes 72% of the bill. Once it is on, cost splits roughly into 32% fresh input, 29% cache reads and 40% output for Opus 5, so output becomes the largest single line.
The most interesting row pair is Opus 5 against GPT-5.6 Sol. Their AA Coding Index values are 78.0 and 77.4, within one point, yet Sol is 2.5× cheaper and completes 882 tasks against 353. Opus 5 does lead on general AA (50.7 vs 47.1) and may justify its premium on hard, open-ended work, but for a budget-constrained agent fleet the data favours Sol. At the bottom, GLM 5.3 Flash completes 10,194 tasks with a Coding Index of 71.5, 6.5 points behind Opus 5. That is 29× the throughput.
Two practical notes. MiniMax M2.7 has a 205K context window, and our 60K average hides long tasks that can exceed it; the other seven list roughly 1M. And Anthropic lists a separate cache-write price ($6.25/M for Opus 5 against $5.00 for ordinary input). If every fresh token is written to the cache, Opus 5's task cost rises 7.9% to $3.06 and the count drops to 327.
6. Workload C: document extraction at volume
| Model | AA | Cost / document | Documents / $1,000 | With 50% batch |
|---|---|---|---|---|
| Claude Opus 5 | 50.7 | $0.1700 | 5,882 | 11,765 |
| GPT-5.6 Sol | 47.1 | $0.0680 | 14,706 | 29,412 |
| Claude Sonnet 5 | 38.4 | $0.0680 | 14,706 | 29,412 |
| Gemini 3.8 Flash | 41.2 | $0.0255 | 39,216 | 78,431 |
| Qwen3.8 Max | 40.3 | $0.0648 | 15,432 | — |
| GLM 5.3 Flash | 41.9 | $0.00490 | 204,082 | — |
| DeepSeek V4 Pro | 30.9 | $0.00660 | 151,559 | — |
| MiniMax M2.7 | 23.2 | $0.00697 | 143,431 | — |
This workload is input-dominated: for Opus 5, 88% of the $0.17 per document is the 30,000 input tokens. So the lever is how much text you send, not how long the answer is. Opus 5 handles 5,882 documents a month; Sol 14,706; Gemini 3.8 Flash 39,216; GLM 5.3 Flash 204,082.
If the job is not latency-sensitive, batch pricing matters. Our database lists separate batch entries for Opus 5, Sol, Sonnet 5 and Gemini 3.8 Flash at exactly 50% of the standard price, which doubles the count to 11,765, 29,412, 29,412 and 78,431. Even batched, Opus 5 handles roughly one seventeenth of what GLM 5.3 Flash does in real time. Batch discounts for the other models are not included, because their batch listings are either missing or not at a consistent discount.
7. Sensitivity: what doubles your bill
The multipliers below are computed from the same formula. Each is how much the monthly cost rises (and units fall) when one assumption changes.
| Model | B: no caching | B: 2× output | A: 3× output (reasoning) | C: 3× output |
|---|---|---|---|---|
| Claude Opus 5 | 3.57× | 1.40× | 1.67× | 1.24× |
| GPT-5.6 Sol | 3.57× | 1.40× | 1.67× | 1.24× |
| Claude Sonnet 5 | 3.57× | 1.40× | 1.67× | 1.24× |
| Gemini 3.8 Flash | 3.57× | 1.40× | 1.67× | 1.24× |
| Qwen3.8 Max | 3.74× | 1.26× | 1.46× | 1.15× |
| GLM 5.3 Flash | 2.98× | 1.23× | 1.50× | 1.16× |
| DeepSeek V4 Pro | 4.67× | 1.22× | 1.33× | 1.10× |
| MiniMax M2.7 | 2.89× | 1.26× | 1.57× | 1.19× |
- No prompt caching. On the agent workload this multiplies cost by about 3.6× for the Anthropic, OpenAI and Google models (3.74× for Qwen) and 4.67× for DeepSeek, whose cache price is 8% of its input price. Of everything here, caching is the biggest lever.
- Output-heavy behaviour. Exposure follows the output-to-input price ratio. Doubling agent output raises Opus 5 by 1.40× but DeepSeek by 1.22×. Three times the output on the document job costs Opus 5 an extra 24% and GLM 5.3 Flash 16%.
- Reasoning tokens. The only reasoning dimension in our data is on Google models, and it equals the output price, so hidden thinking tokens are simply output tokens. If thinking triples billed output on the chatbot, cost rises 1.67× on Opus 5 and 1.50× on GLM 5.3 Flash. Check your provider usage logs for the real output count.
- Price state of the budget models. DeepSeek V4 Pro shows what a router flap does when nothing else changes:
| DeepSeek V4 Pro price state | Input / Output $/M | A: conversations | C: documents |
|---|---|---|---|
| Current (2026-10-05) | $0.2088 / $0.4176 | 266,071 | 151,559 |
| Early price (2026-05-03 to 08-10) | $0.435 / $0.87 | 127,714 | 72,748 |
| 30-day high (2026-09-12 to 09-18) | $1.60 / $3.20 | 34,722 | 19,778 |
Between today's price and the 30-day high, the same budget buys 7.7× fewer conversations. Budget-tier plans should carry a margin for that, or pin a specific provider. Not modelled: tokenizer differences (the same text counts differently across models), retries, rate limits, volume discounts and tool-call overhead beyond tokens.
8. Verdict and how to adapt this
Support chatbot. Unit price is not your constraint: every model here handles more than 8,000 conversations for $1,000. A common pattern is routing: send 90% of conversations to GLM 5.3 Flash and escalate 10% to Opus 5, which costs 0.9 × $0.0030 + 0.1 × $0.1125 = $0.01395 per conversation and covers about 71,685 conversations. That is an illustration of the arithmetic, not a tested routing policy.
Coding agent. Turn on prompt caching first, since it is worth more than any model switch. Then compare GPT-5.6 Sol (882 tasks, Coding 77.4) with Opus 5 (353 tasks, Coding 78.0) on your own repository before paying 2.5× more.
Document extraction. Trim input, use batch pricing where you can wait, and test a budget model first. Move up a tier only if your own evaluation shows the quality gap matters.
Before you commit. The quality columns are averages over general benchmarks. For support, the number that matters is resolution rate and escalation rate on your own transcripts; for agents, it is the share of tasks that finish without a human fix; for extraction, it is field-level accuracy against a labelled sample. Run 100 real examples through two candidate models, measure that number, and divide the cost per unit by it. A model that is 10× cheaper per unit but needs three times the retries is only 3× cheaper.
To plug in your own token counts, use the cost calculator. For the broader methodology see our token pricing guide and the value benchmark.