Low-cost API for frontier open-source models

Kimi K2.6 and MiniMax M2.7 perform at the level of GPT and Claude in many benchmarks. Each model costs half of its vendor's own price — about 90% less than flagship APIs. OpenAI-compatible: switching takes two lines of code.

Try it for free See the benchmarks

Every new account includes a free trial credit · prepaid after that, no subscription · $5 minimum top-up ≈ 5–22M tokens

You pay Khidi (kimi-k2.6)

$25

33M tokens / mo — drag me

Claude Opus 4.8 charges

$278

−89%your bill vs Opus 4.8 +1010%the Opus premium, same tokens

Models & pricing

Kimi K2.6, MiniMax M2.7, GLM-5.2 — half their vendors' prices.

Same weights, same capabilities — each priced at 50% of what its own maker charges. Input and output rates per 1M tokens, metered per request. Prepaid: $5 minimum top-up, no subscription, usage stops at your budget.

kimi-k2.6

Kimi K2.6 agentic workhorse

Beats the flagships on tool-use and research benchmarks (HLE w/ tools, DeepSearchQA). 256K context. The all-rounder. Full benchmarks →

$0.50 in · $2.00 out
vendor: $0.95 / $4.00

minimax-m2.7

MiniMax M2.7 high-volume value

10B active parameters — fast and remarkably capable for the price. Best for bulk pipelines, classification, drafts. Full benchmarks →

$0.15 in · $0.60 out
vendor: $0.30 / $1.20

glm-5.2

GLM-5.2 coming soon

The newest open flagship — top open model on the Artificial Analysis Index, 1M-token context. Arriving at Khidi shortly. Full benchmarks →

$0.70 in · $2.20 out
vendor: $1.40 / $4.40

Published benchmarks · exact versions named · Apr–Jun 2026

Open models stopped playing catch-up.
On real work they match the flagships — and often beat them.

These are the three models we serve, against the flagships you're paying 15–80× more for — including the newest, Claude Opus 4.8 and Claude Fable 5. Where they win, we show it. The margins are the story. Sources under each chart.

SWE-Bench Pro GLM-5.2 leads; Kimi ties GPT

Resolving real GitHub issues — the hardest, contamination-resistant software-engineering benchmark.

GLM-5.2

62.1

Kimi K2.6

58.6

GPT-5.5

58.6

GPT-5.4 (xhigh)

57.7

MiniMax M2.7

56.2

Gemini 3.1 Pro

54.2

Claude Opus 4.6

53.4

Kimi/GPT-5.4/Gemini/Opus 4.6: Moonshot AI release, Apr 2026. GLM-5.2 vs GPT-5.5: public compilations, Jun 2026. MiniMax M2.7: vendor-reported, Jun 2026. Claude Opus 4.8 and Fable 5 have no directly comparable published runs on this benchmark (see FrontierSWE and Code Arena below for newest-flagship comparisons). Highlighted bars = models served by Khidi.

FrontierSWE 0.7 points from the flagship

Frontier-grade software engineering. The gap between the best closed model and open weights is now under one point.

Claude Opus 4.8

75.1

GLM-5.2

74.4

GPT-5.5

72.6

Official scorecard readings, Jun 2026. Price difference vs Opus 4.8: ~80×. GLM-5.2 is also the top open-weight model on the Artificial Analysis Intelligence Index (51 — 4th overall, behind only Fable 5, Opus 4.8 and GPT-5.5). Kimi K2.6 and MiniMax M2.7 have no published FrontierSWE runs.

Humanity's Last Exam — with tools only the newest flagship stays ahead

The hardest knowledge benchmark, solved the way real agents work: with search and tools.

Claude Opus 4.8

57.9

GLM-5.2

54.7

Kimi K2.6

54.0

Claude Opus 4.6

53.0

GPT-5.5

52.2

GPT-5.4 (xhigh)

52.1

Gemini 3.1 Pro

51.4

Sources: Moonshot AI K2.6 release (Apr 2026); GLM-5.2/GPT-5.5/Opus 4.8 from public compilations (Jun 2026). The newest flagship leads by 3.2 points — at ~80× the price. MiniMax M2.7 has no published HLE-with-tools run.

Code Arena: Frontend above every Opus variant

Head-to-head Elo on real web-development tasks, judged by humans. Only one model in the world scores higher.

Claude Fable 5

1665

GLM-5.2 (Max)

1595

best Claude Opus variant

<1595

Source: Arena.ai leaderboard, 25 Jun 2026 (bar length ∝ Elo). GLM-5.2 beat every Claude Opus variant in matched pairings, incl. 55.0% win share vs Opus 4.7 (Thinking). GLM-5.2 is also the top open-weight model on the Artificial Analysis Intelligence Index — 4th overall, behind only Fable 5, Opus 4.8 and GPT-5.5.

DeepSearchQA (f1) +13.9 points over GPT-5.4

Deep factual retrieval and synthesis across long contexts — the core skill of research agents.

Kimi K2.6

92.5

Claude Opus 4.6

91.3

Gemini 3.1 Pro

81.9

GPT-5.4 (xhigh)

78.6

Source: Moonshot AI K2.6 release, Apr 2026; flagship scores cited there from the Claude Opus 4.7 System Card. GLM-5.2 and MiniMax M2.7 have no published DeepSearchQA runs.

SWE-Bench Verified statistical tie with the flagship

Single-file bug fixes on real repositories. The top models are now within one point of each other.

Claude Opus 4.6

80.8

Gemini 3.1 Pro

80.6

Kimi K2.6

80.2

MiniMax M2.7

78.0

Spread across the top three: 0.6 points. Price difference: ~80×. Sources: Moonshot AI (Apr 2026), MiniMax vendor-reported (Jun 2026).

The economics

Same work. Two invoices.

A typical coding-agent workload: 50M input + 10M output tokens per month.

Monthly bill comparison

GPT-5.5 ($5 / $30)$550.00

Claude Opus 4.8 ($5 / $25)$500.00

GLM-5.2 at its vendor ($1.40 / $4.40)$114.00

Kimi K2.6 at its vendor ($0.95 / $4.00)$87.50

Khidi — kimi-k2.6 ($0.50 / $2.00)$45.00

Khidi — minimax-m2.7 ($0.15 / $0.60)$13.50

Khidi — glm-5.2 ($0.70 / $2.20) — coming soon$57.00

Quickstart

Two lines. That's the migration.

Get a key

Sign up, click "Create my API key", done — with a free trial budget included. Top up later in USD or USDT. No subscription, no sales call.

Change two lines

client = OpenAI(
  base_url="https://api.khidi.ai/v1",
  api_key="sk-…",
)

client.chat.completions.create(
  model="kimi-k2.6",  # per request
  messages=[…],
)

Ship

Pick the model per request — swap by changing one string: kimi-k2.6 or minimax-m2.7 (glm-5.2 coming soon). In no-code tools (Cursor, LibreChat, n8n) just type the model ID into the "custom model" field. Streaming, tool calling and long context out of the box.

Prefer Claude Code or the Anthropic SDK? also compatible

Khidi speaks the Anthropic Messages format too — run open models inside Claude Code, at a fraction of the price.

export ANTHROPIC_BASE_URL="https://api.khidi.ai"
export ANTHROPIC_AUTH_TOKEN="sk-…"
claude --model kimi-k2.6

Free trial budget · prepaid · no subscription

Get your API key in two minutes

Sign up, create your key, start calling — it comes with a small trial budget so you can test before paying a cent. Then prepaid top-ups with a budget that stops exactly where you set it.

We never store prompt content or use it for training — see the Privacy Policy for how prompts are processed. Prepaid budgets: your key stops when the budget is spent — no surprise bills.

Contact

Questions? Talk to us.

Every message lands in a human inbox — we reply within one business day.

Email us directly

We reply within one business day.

General & technical — [email protected] Billing & refunds — [email protected]

Kimi K2.6, MiniMax M2.7, GLM-5.2 — half their vendors' prices.

Kimi K2.6 agentic workhorse

MiniMax M2.7 high-volume value

GLM-5.2 coming soon

Open models stopped playing catch-up.On real work they match the flagships — and often beat them.

SWE-Bench Pro GLM-5.2 leads; Kimi ties GPT

FrontierSWE 0.7 points from the flagship

Humanity's Last Exam — with tools only the newest flagship stays ahead

Code Arena: Frontend above every Opus variant

DeepSearchQA (f1) +13.9 points over GPT-5.4

SWE-Bench Verified statistical tie with the flagship

Same work. Two invoices.

Monthly bill comparison

Two lines. That's the migration.

Get a key

Change two lines

Ship

Prefer Claude Code or the Anthropic SDK? also compatible

Get your API key in two minutes

Questions? Talk to us.

Email us directly

Open models stopped playing catch-up.
On real work they match the flagships — and often beat them.