Chat Picker

ChatGPT Alternatives: Chinese AI Models and Their Published Pricing

Compares published per-token API prices from Chinese AI providers and shows readers how to test model fit, cost, and source checking in an afternoon.

Sources checked 3 Oct 2026

Chinese-developed API products from DeepSeek, Alibaba Cloud Model Studio, Moonshot Kimi, and Z.AI are possible ChatGPT alternatives, but their vendor pages settle prices and model terms—not performance on your files. As read on October 1, 2026, the DeepSeek Models & Pricing page, Alibaba Cloud supported-models page, Kimi API pricing page, and Z.AI pricing overview document different combinations of token rates, context, caching, and output limits. Chat Picker has not tested the services; use the plan below to test them yourself.

Know these limits before you start

Start with billing mechanics. The DeepSeek Models & Pricing page, read October 1, 2026, applies peak rates from 01:00–04:00 and 06:00–10:00 UTC on weekdays except Chinese public holidays. All other hours are off-peak and cost half as much. Both listed models support thinking and non-thinking modes, with thinking as the default. DeepSeek-Flash supports vision; DeepSeek-V4-Pro does not.

These selected published rows show why the billing unit matters:

Provider and model Published rate, USD per 1M tokens Window or limit
deepseek-flash (DeepSeek-V4.1-Flash), DeepSeek pricing, read 2026-10-01 Cache-miss input: $0.15 off-peak, $0.30 peak; cache-hit input: $0.003, $0.006; output: $0.6, $1.2 1M context; 384K maximum output
deepseek-v4-pro (DeepSeek-V4-Pro-0813), DeepSeek pricing, read 2026-10-01 Cache-miss input: $0.66 off-peak, $1.32 peak; cache-hit input: $0.022, $0.044; output: $1.98, $3.96 1M context; 384K maximum output
kimi-k3, Kimi API pricing, read 2026-10-01 Five-minute cache write: $3; one-hour cache write: $6; cached input: $0.30; input: $3; output: $15 1,048,576-token context
kimi-k2.7-code, Kimi API pricing, read 2026-10-01 Cache-hit input: $0.19; cache-miss input: $0.95; output: $4 262,144-token context
GLM-5.3-Flash, Z.AI pricing, read 2026-10-01 Input: $0.15; cached input: $0.03; output: $0.50 No context figure given on this page
GLM-4.6V, Z.AI pricing, read 2026-10-01 Input: $0.30; cached input: $0.05; output: $0.9 No context figure given on this page

Alibaba Cloud’s supported-models page, read October 1, 2026, directs audio or video analysis and text generation to qwen3.8-omni-flash, but that page publishes no Qwen token price. The Z.AI pricing page, read on the same date, also marks GLM-4.7-Flash and GLM-4.6V-Flash as Free; treat that as the listed rate, not a forward promise. The Kimi pricing page says its listed prices exclude applicable taxes, calculated at checkout for your jurisdiction.

Those four API pages do not settle whether submitted inputs are used for training or how long they are retained. Check the terms attached to your account before using sensitive material. If you include ChatGPT as a baseline, OpenAI’s accuracy note, read October 1, 2026, says it can produce incorrect or misleading output and may sound confident when wrong; verify important information.

The test plan

Use a fresh session for each run. Keep the plan, exact model identifier, mode, and settings fixed so that one visible change does not mix model and configuration effects.

  1. Instruction following. Prepare a 400-word non-sensitive operations memo containing names, dates, four numerical constraints, and one deliberate conflict. Give: “Rewrite this memo for a US audience in exactly five bullets. Keep every number, date, and name. If two instructions conflict, add one final line beginning ‘Needs review:’ and explain the conflict.” Check the bullet count, preserved values, and conflict disclosure. Log each requirement as pass or fail and record unsupported changes.

  2. Structured output. Prepare a 12-row CSV with customer IDs, dates, currencies, and positive or negative totals. Give: “Return valid JSON only, with one object per row and the keys customer_id, total, currency, and valid. Set valid to false when total is negative. Do not add commentary.” Check JSON validity, row order, data types, and signs. Log the raw response, parser errors, and input and output token counts.

  3. Code debugging. Give the model this function: def average(values): return sum(values) / len(values). Add: “Find one input that can make this function fail. Return a valid JSON object with keys input, actual, and expected; describe the error in actual and the desired behavior in expected.” Check the arithmetic and whether your test reproduces the stated error. Log the response, verification steps, and any proposed correction.

  4. Long-context retrieval. Prepare a plain-text document with numbered paragraphs. Put “The review window closes on 18 December 2026.” in one paragraph and “No external archive is required.” in another. Give: “Find the paragraph containing ‘review window’ and the paragraph containing ‘external archive.’ For each, return its paragraph ID and exact sentence. If either phrase is absent, write Not found.” Check the IDs, quotations, and omissions. Log any truncation notice, token count, and retrieval failures.

  5. Documented modalities or modes. For a model whose vendor page documents image input, prepare a chart with five labeled fictional values. Give: “Read the five labeled values and return JSON with keys label_1 through label_5. Preserve units exactly and use null for an unreadable label.” For other models, mark this test not applicable. If testing DeepSeek modes, change only the thinking setting and record it. Log the exact model, mode, settings, and result.

How to read your results

Set must-pass conditions before testing: valid requested format, unchanged critical values, traceable source text, acceptable data terms, and a charge within your budget. Classify each run against those conditions, then review the vendor’s actual token usage and operational notes. Do not let a low charge compensate for a failed privacy or validity condition.

Calculate cost from recorded usage:

Cost = cache-miss input ÷ 1,000,000 × input rate + cache-hit input ÷ 1,000,000 × cached-input rate + output ÷ 1,000,000 × output rate + cache-write charges.

Use DeepSeek’s applicable peak or off-peak rate and include Kimi K3’s selected cache-write charge. Keep every attempt in the log. If several options pass, choose according to the workload you will repeat, the costs you will actually incur, and the constraints that must remain intact.

Where the plans differ

The Claude vs ChatGPT vs Gemini comparison matrix separates the Free plan, Low-cost tier, Main paid plan, Context window in app, How usage limits are described, Training on your chats, and Ads rows. Those rows help you check access, billing, limits, data settings, and advertising, but they do not supply quality results. Its consumer subscription prices also should not be treated as API token prices for the Chinese-developed services above.

Chat Picker records the vendor source and reading date for each comparison figure, marks unconfirmed entries “Not verified,” and leaves them blank. An older version of this page contained statistics and test results without sources; they have been removed.

Sources

For current prices and limits, see the dated comparison pages.

Compare assistants