Baichuan AI pricing

About

Baichuan AI (百川智能 / Baichuan Intelligence) is a Beijing-based foundation-model lab founded on March 24, 2023 by Wang Xiaochuan (王小川), the former CEO of Chinese search engine Sogou, with co-founder Ru Liyun. Its core team was assembled from Sogou, Baidu, Huawei, Microsoft, ByteDance, and Tencent. One of China’s so-called “AI tigers,” Baichuan moved fast: less than 100 days after founding it released the open-weight Baichuan-7B and Baichuan-13B Chinese models — free to download and commercially licensable — which together passed one million downloads.

Baichuan raised roughly $1 billion across four rounds, including a ~$300M round in October 2023 from Alibaba, Tencent, and Xiaomi, reaching a valuation around $2.75–2.8 billion by mid-2024. It opened a paid developer API on May 22, 2024 alongside its first consumer assistant, “Baixiaoying” (百小应), and was a participant in the brutal mid-2024 Chinese token price war, where rivals cut API rates by as much as 97%.

What makes Baichuan distinctive in this cluster is the strategic narrowing. In March 2025, the company pivoted hard to healthcare — it dropped its three-month-old Baichuan4-Finance vertical model, cut financial-sales staff, and refocused entirely on medical AI. Today its public face is “百小医” (Bai Xiao Yi), an “AI family doctor,” backed by the M-series medical models (M1, M2, M3, M4) that headline HealthBench leadership claims. The pricing structure reflects both eras: a still-live RMB per-token API across general Baichuan3/4 models, and a deliberately free medical base (Baichuan-M3-Plus via the 海纳百川 program) aimed at owning the clinical ecosystem rather than monetizing inference directly.

For current rates, see the Baichuan API price card.

Pricing summary : RMB per-token API plus free medical base and open weights

Baichuan AI runs an RMB-native pure usage API layered over a freemium open-weight and medical strategy. The dimensions are:

• General models (blended per-1K-token rate) — most general models bill a single combined input+output rate: Baichuan4 at ¥0.1/1K, Baichuan4-Turbo at ¥0.015/1K, Baichuan4-Air at ¥0.00098/1K, Baichuan3-Turbo at ¥0.012/1K, Baichuan3-Turbo-128k at ¥0.024/1K, and Baichuan2-Turbo at ¥0.008/1K. Baichuan2-53B uses a time-varying off-peak rate (¥0.01/1K overnight vs ¥0.02/1K daytime).
• Medical models (split input/output) — Baichuan-M3 at ¥0.01 in / ¥0.03 out per 1K, the evidence-anchored M3-Plus at ¥0.005 in / ¥0.009 out per 1K, M2 at ¥0.002 in / ¥0.02 out per 1K, M2-Plus at ¥0.01 in / ¥0.03 out per 1K.
• Tool calls — medical search and web search are each ¥0.03 per call; the Assistants API is currently free.
• Embeddings — Baichuan-Text-Embedding at ¥0.0005/1K tokens.
• Free tiers — the 海纳百川 program gives the M3-Plus API away free to qualifying medical institutions, and open-weight Baichuan/Baichuan2/M-series models are free to download and commercially license by application.

What makes this different: Baichuan prices in RMB per 1,000 tokens (not per-million USD like Western peers), the general flagship-to-Air spread is roughly 100× within one family, and the company has deliberately set its strategic medical base to zero — monetizing distribution and ecosystem position rather than inference. USD conversions throughout are approximate at ~¥7.2/$.

Pricing by product

General models — API (RMB ¥ per 1,000 tokens, blended input+output)

Medical models — API (RMB ¥ per 1,000 tokens, split input / output)

Specialized & free

Sales motions across products: self-serve / PLG for the public API and open weights; an application-gated free program (海纳百川) and partnership/sales-led motion for clinical and hospital deployments.

Hidden costs : What Baichuan AI users actually pay

Baichuan’s headline rates are unusually low, but three things shape a real bill: the split input/output rates on medical models (output is 3–10× input), the per-call search fees layered on top of tokens, and the fact that the cheapest general model (Air) and the strategic medical base (M3-Plus, free via 海纳百川) are effectively loss leaders — so most real spend lands on the mid-tier models. Two archetypes show how the total assembles. USD figures are approximate at ~¥7.2/$.

Archetype 1 — a developer running a general assistant on Baichuan4-Turbo. Assume ~50M input + ~15M output tokens/month (blended billing at ¥0.015/1K), plus 20,000 web-search tool calls.

The lesson: because Baichuan4-Turbo bills a single blended rate, you can’t shift cost by trimming output, and the per-call search fee (¥0.03) quietly rivals the token line once an assistant searches the web on most turns. Dropping to Baichuan4-Air (¥0.00098/1K) cuts the token line ~15× — but you trade flagship quality.

Archetype 2 — a clinical-decision-support app on the medical M3 line. Assume ~30M input + ~10M output tokens/month on Baichuan-M3 (split rates), with auto medical search on ~100,000 turns.

Here the surprise is the medical-search fee dominating the bill — and the strategic out is to qualify for the 海纳百川 program, which provides the M3-Plus base API for free, leaving only the per-call search and any tooling as variable cost. The free base is the whole point: Baichuan wants the clinical workflow, not the inference margin.

Want to estimate your own Baichuan AI bill? Use the Baichuan AI pricing calculator to model your costs based on token volume, model choice, and search-call frequency.

Pricing evolution : Baichuan AI pricing history and changes

Baichuan’s pricing arc tracks its strategic one: open weights first (mid-2023), a paid RMB per-token API at the May 2024 Baichuan4 launch, compression in the mid-2024 price war (Baichuan4-Air at near-zero by November), then a March 2025 narrowing to healthcare that re-centered the price sheet on free and cheap medical models. The dated milestones below are reconstructed from primary announcements and contemporaneous press.

Cadence

Tracked range: 2023 Q2–2026 Q2. Quarters not listed had no publicly announced price or SKU change material to this corpus.

Notable changes

• 2023-06 — Baichuan-7B / 13B open-weight models released free; >1M downloads, seeding the developer ecosystem before any meter (HuggingFace).
• 2024-05-22 — Baichuan4 (¥0.1/1K) + Baixiaoying assistant launch; four APIs open (Baichuan4, Baichuan3-Turbo, Baichuan3-Turbo-128k, Assistant). Claimed China #1 SuperCLUE (reported by SCMP, BrainTitan).
• 2024-06 — China “AI tiger” token price war: Alibaba cut up to 97%, peers followed; Baichuan a participant. Wang Xiaochuan framed it as cloud vendors “opening a new battlefield” (reported by Global Times, Disruption Banking).
• 2024-11 — Baichuan4-Turbo (¥0.015/1K, ~80% of GPT-4o price) and MoE Baichuan4-Air (¥0.00098/1K, near-zero) added; long context to ~192K tokens.
• 2024-12 — Baichuan4-Finance launched, claimed to beat GPT-4o on a dozen-plus financial tasks (reported by TMTPost).
• 2025-03 — Healthcare pivot: Baichuan4-Finance abandoned ~3 months post-launch, financial-sales staff cut, two co-founders departed; focus shifts to medical M-series and the 百小医 family-doctor product (reported by The Wire China, 2025-03-30).

The finance-to-healthcare pivot in detail

The March 2025 pivot is the defining lifecycle event. Baichuan had spent late 2024 building enterprise breadth — a general assistant (Baixiaoying), a long-context model, and a dedicated Baichuan4-Finance vertical it claimed surpassed GPT-4o on financial tasks. Within roughly three months it reversed: financial-sales employees were told to resign, the finance model was shelved, and two co-founders left to start their own venture. The replacement strategy was a single defensible vertical — medical AI — anchored by HealthBench-leading open M-series models, a consumer “AI family doctor” (百小医), and the 海纳百川 program that gives the M3-Plus base API away free to institutions serving healthcare workers. The pricing implication is stark: rather than competing on general-purpose token price (a war it could not win against Alibaba and DeepSeek), Baichuan set its strategic base to zero and bet on owning the clinical distribution layer.

What’s unique : Baichuan AI’s distinctive pricing mechanics

1. RMB-native, blended-rate token pricing. Unlike Western peers who quote per-million-token USD with separate input/output, most Baichuan general models bill a single combined ¥/1,000-token rate — Baichuan4 ¥0.1, Turbo ¥0.015, Air ¥0.00098. The blended rate simplifies the headline but removes the lever of trimming output to cut cost, and the ~100× spread from flagship to Air inside one family lets developers trade quality for price within the same SDK call.

2. A strategic base priced at zero. The 海纳百川 (“hundred rivers into the sea”) program gives the evidence-anchored Baichuan-M3-Plus medical API away free, permanently, to qualifying institutions. This is not a trial — it’s a deliberate decision to monetize ecosystem position and downstream clinical workflows rather than inference. Pricing the base at zero is a distribution strategy, not a discount.

3. Open weights given away to seed a paid clinical layer. Baichuan releases its medical models (M1 14B, M2 32B, M3 235B) as open weights, free for commercial use by license application — including HealthBench-leading models that run on a single RTX 4090. The free weights build credibility and adoption; the hosted API, evidence-anchoring, and medical search are the priced layer on top — a hedge against lock-in that closed-weight rivals can’t match.

Strengths & weaknesses

Billing UX : usage tracking and overage controls

• Developer platform console — API keys, usage, and billing are managed through platform.baichuan-ai.com; the public price card lists every model’s ¥/1K rate and context window.
• Blended vs split metering — most general models meter a single combined input+output rate, while medical models meter input and output separately; the console attributes cost per model.
• Per-call tool charges — medical search and web search are billed at ¥0.03/call on top of tokens, surfaced as separate line items rather than bundled into the token rate.
• Off-peak rate window — Baichuan2-53B charges ¥0.01/1K overnight (00:00–08:00) vs ¥0.02/1K daytime, a time-of-day control that rewards scheduling batch work overnight.
• Free-program gating — 海纳百川 access requires application + Baichuan approval, mandatory “Powered by 百川” attribution, and is restricted to clinical decision support and medical education (no data-production use).
• Open-weight self-host — teams that prefer to avoid the meter entirely can download M-series weights and run inference themselves, applying for a commercial license.

Strategic wins : Why Baichuan AI’s pricing decisions worked

1. Setting the strategic base to zero

By giving the Baichuan-M3-Plus medical API away free via 海纳百川, Baichuan converted its core asset into a distribution wedge into hospitals and clinical tools. It monetizes the ecosystem and downstream services rather than the inference, a bet that owning the clinical workflow is worth more than per-token margin. See usage-based pricing strategy for why metering the surrounding service can beat metering the artifact.

2. Open weights as credibility, not commodity

Releasing HealthBench-leading M-series models as free, commercially-usable open weights — some running on a single consumer GPU — built developer trust and clinical adoption that a closed model couldn’t buy. The free weights are marketing; the hosted, evidence-anchored API is the revenue line. This mirrors the shift away from rigid licensing toward open-core monetization.

3. Narrowing to a defensible vertical

Rather than keep losing a general-purpose price war against Alibaba and DeepSeek, Baichuan picked one vertical — healthcare — where evidence-anchoring, low hallucination, and HealthBench leadership are a real moat. Concentrating its pricing and product story on medical AI is a focus bet that trades breadth for defensibility. See choosing the right usage metric for why a narrow, value-aligned surface can outprice a broad one.

Areas to improve : Gaps in Baichuan AI’s pricing approach

1. Publish an international / USD price card

The price card is RMB-native and oriented to China. Developers outside China face an unpredictable conversion step and no first-class USD card, which limits Baichuan’s reach versus peers who publish USD per-million-token rates. A clear international card would lower the evaluation barrier.

2. Surface per-call search costs in the headline

Medical and web search are ¥0.03/call on top of tokens and, on search-heavy clinical workloads, can exceed the token bill. Burying them below the model table makes total cost hard to predict. A combined “estimated cost per query” view would make the real bill legible before a customer commits.

3. Stabilize the SKU roadmap

Launching and abandoning Baichuan4-Finance within roughly three months — and de-emphasizing general models after the healthcare pivot — signals roadmap volatility that makes long-term pricing commitments risky for buyers. Clearer signals about which surfaces are durable would help mid-market customers plan. Compare how other AI companies stage enterprise transparency.

Key takeaways

1. Price the strategic base at zero to win distribution. Giving the M3-Plus medical API away free via 海纳百川 trades inference margin for clinical-ecosystem position — a bet that owning the workflow beats charging for tokens.
2. Open weights build the moat; the API sells the service. Free, commercially-usable medical models seed adoption and credibility, while evidence-anchoring and hosted inference are the priced layer on top.
3. A blended token rate is simpler but less steerable. Combining input+output into one ¥/1K rate is easy to read, but it removes the customer’s lever to optimize cost by trimming output — a trade-off worth making deliberately.
4. Narrowing can be a pricing decision. Exiting a general-purpose price war to focus on one defensible vertical re-centers the whole price sheet on free and cheap medical models — focus as strategy.
5. Currency and geo shape reach. RMB-native pricing keeps Baichuan legible at home but raises friction for international developers — a reminder that the currency you quote in is itself a go-to-market choice.

UBP implications

1. The cheapest viable price is sometimes zero. When a vertical’s value is in the workflow rather than the inference, setting the base model to free (as 海纳百川 does for M3-Plus) can be the rational monetization move — UBP practitioners should ask whether the meter belongs on the model or the surrounding service.
2. Open weights reframe the priced unit. When the model is free to download, the priced dimension becomes hosted inference, evidence-anchoring, and search — not the model itself. UBP design has to follow the cost driver customers can’t trivially replicate.
3. Price wars compress the meter toward zero — focus is the escape. Baichuan4-Air’s near-zero ¥0.00098/1K shows where commodity token pricing trends; Baichuan’s answer was to narrow to a vertical where a richer value metric (clinical outcomes, evidence accuracy) can re-anchor price away from raw tokens.

Sources

• Baichuan API price card — platform.baichuan-ai.com/prices (accessed 2026-06-11)
• Baichuan AI homepage — healthcare / 百小医 / 海纳百川 (accessed 2026-06-11)
• The Wire China — “Baichuan’s Big Pivot” (accessed 2026-06-11)
• SCMP — Baichuan debuts AI assistant powered by upgraded LLM (accessed 2026-06-11)
• Turing Post — Baichuan Intelligence profile (accessed 2026-06-11)
• Baichuan on HuggingFace (open weights) (accessed 2026-06-11)
• Browse the pricing blueprint corpus

Bottom line

Baichuan AI prices an RMB-native per-1,000-token API spanning a ~100× range — from Baichuan4 at ¥0.1/1K down to the near-zero Baichuan4-Air at ¥0.00098/1K, a relic of China’s 2024 token price war. But the real story is the March 2025 pivot to healthcare: Baichuan dropped its finance model, narrowed to HealthBench-leading medical M-series models, and set its strategic Baichuan-M3-Plus base to free via the 海纳百川 program — betting that owning the clinical workflow beats winning a price war on tokens. The main friction is RMB-native, China-geo pricing and per-call search fees that can dominate a search-heavy bill.

Want to compare Baichuan AI against other foundation-model providers? See Mistral AI, or browse the full pricing blueprint.

Pricing timeline : Major events on a vertical axis

Each milestone below corresponds to a public pricing change, product launch, or material adjustment. Major events use a filled marker; minor adjustments use a faded one.

Live snapshot: medical M-series pricing + free 海纳百川 base

Jun 2026

Captured live RMB price card: general models Baichuan4 ¥0.1/1K, Baichuan4-Turbo ¥0.015/1K, Baichuan4-Air ¥0.00098/1K; medical Baichuan-M3 ¥0.01 in / ¥0.03 out per 1K and M3-Plus ¥0.005 in / ¥0.009 out per 1K. The 海纳百川 program gives the M3-Plus API away free to qualifying medical institutions; open M-series weights remain free to commercialize.

Healthcare pivot — finance dropped, medical M-series focus

Mar 2025

Baichuan narrowed hard to healthcare: financial-sales staff were told to resign, Baichuan4-Finance was abandoned ~3 months post-launch, and two co-founders left. Focus shifted to medical AI — the 'AI pediatrician' / 百小医 family-doctor product and the M-series (M1/M2/M3/M4) medical models. (Source: The Wire China, 2025-03-30.)

Baichuan4-Finance launched (later abandoned)

Dec 2024

Baichuan released Baichuan4-Finance, claiming it surpassed OpenAI's 4o on a dozen-plus financial tasks. This vertical SKU was dropped roughly three months later in the March 2025 pivot — a short-lived enterprise bet. (Source: The Wire China, 2025-03.)

Baichuan4-Turbo + Baichuan4-Air added at the cheap end

Nov 2024

Baichuan released Baichuan4-Turbo (¥0.015/1K, pitched at ~80% of GPT-4o's price for enterprise) and the MoE Baichuan4-Air (¥0.00098/1K — near-zero), emphasizing cost-efficiency and long context (to ~192K tokens in API). The Air rate is the clearest artifact of the price war's race to the bottom.

China 'AI tiger' token price war — rates slashed

Jun 2024

Baichuan was a participant in the mid-2024 Chinese large-model price war, where Alibaba cut API prices up to 97% and Baidu and others followed, some toward near-zero. CEO Wang Xiaochuan framed the cuts as cloud vendors opening a new battlefield. Baichuan's blended rates compressed sharply in this period. (Source: Global Times, Disruption Banking, 2024.)

Baichuan4 + Baixiaoying assistant launch; first paid APIs open

May 2024

Baichuan released its Baichuan4 flagship and first AI assistant 'Baixiaoying' (百小应), opening four APIs: Baichuan4, Baichuan3-Turbo, Baichuan3-Turbo-128k, and the Assistant API. This established the RMB per-1,000-token meter. Baichuan4 launched at ¥0.1/1K, claiming China's #1 SuperCLUE score. (Source: SCMP, BrainTitan, 2024-05.)

Baichuan-7B / Baichuan-13B open-weight models released free

Jun 2023

Less than 100 days after its March 2023 founding, Baichuan released the Baichuan-7B and Baichuan-13B open-weight Chinese LLMs — free to download and commercially licensable. Downloads surpassed one million. Monetization was deferred; the open models seeded a developer ecosystem before any API meter existed.