Developer API cost planner

Open-Weight Model Cost Calculator

Turn token usage into a practical monthly API budget. Compare six popular open-weight models, test caching assumptions, account for retries and provider markup, and find a better fit for your application.

Calculate my cost View price table

Pricing snapshot: August 14, 2026 · USD per 1 million tokens

Cost spectrum1M input + 1M output

DeepSeek V4 Flash 0731

$0.420

DeepSeek V4 Pro

$1.305

Kimi K2.7 Code

$4.950

GLM-5.3

$5.800

GLM-5.2

$5.800

Kimi K3

$18.000

GLM-5.3 uses a GLM-5.2 planning proxy; taxes and gateway fees excluded

Estimate traffic

Use requests per day and active days to model an average month.

Measure tokens

Enter average input and output tokens for one completed request.

Model reality

Add cache hits, retries, tool calls and any gateway markup.

Compare budget fit

Rank models with published rates using the same workload assumptions.

01 / Workload calculator

Build a realistic monthly estimate

All calculations run in your browser. No usage information is submitted or stored.

Best projected fit—

Monthly tokens—

Effective cache—

Models in budget—

02 / Price reference

Direct API pricing at a glance

Cache-miss input, cache-hit input and generated output are billed separately. Lower token price does not guarantee the best result for every workload.

Model	Provider	Input	Cached input	Output	1M in + 1M out	Position
GLM-5.3	Z.ai	~$1.400000	~$0.260000	~$4.40	~$5.800	Planning proxy · New coding frontier
Kimi K3	Moonshot AI	$3.000000	$0.300000	$15.00	$18.000	Premium capability
GLM-5.2	Z.ai	$1.400000	$0.260000	$4.40	$5.800	Balanced agents
DeepSeek V4 Pro	DeepSeek	$0.435000	$0.003625	$0.87	$1.305	Low-cost reasoning
Kimi K2.7 Code	Moonshot AI	$0.950000	$0.190000	$4.00	$4.950	Coding specialist
DeepSeek V4 Flash 0731	DeepSeek	$0.140000	$0.002800	$0.28	$0.420	Lowest current cost

Values marked with “~” are planning estimates: GLM-5.3 currently uses GLM-5.2’s published $1.40 input, $0.26 cached-input and $4.40 output rates until Z.ai publishes a dedicated direct API rate. Rates exclude taxes, currency conversion, minimum charges, tool fees, storage, embeddings, web search and third-party gateway markup. “Cached input” applies only when the provider reports a qualifying cache hit.

03 / New release

GLM-5.3: stronger coding from post-training

Z.ai released GLM-5.3 on August 14, 2026. It uses the same base model as GLM-5.2, with the reported gains coming from scaled post-training on broader, longer and more realistic task environments.

Provisional calculator rate$1.40 in · $4.40 out

The calculator uses GLM-5.2’s published rate as a planning proxy, including $0.26 per million cached-input tokens. This is not presented as an official GLM-5.3 rate and should be replaced when Z.ai publishes one.

GLM-5.3 is available to Coding Plan users through points-based quotas. Off-peak calls consume 50% of standard points outside Monday–Friday, 14:00–18:00 UTC+8.

Code Bench gain+50%

Z.ai reports a 50% improvement over GLM-5.2 on its private coding-agent benchmark.

Terminal Bench 3.028.3

Up from 4.6 for GLM-5.2 in Z.ai’s published evaluation.

DeepSWE v1.166.9

Up from 46.2, with long-horizon software-engineering emphasis.

CyberGym84.5%

Z.ai reports a benchmark-leading result, up from GLM-5.2’s 77.2%.

ExploitBench54.4%

More than double GLM-5.2’s reported 24.4% result.

Token efficiency~75K

At Max effort, 34.5% on Z.ai Code Bench at roughly 75K output tokens, versus 23.4% at 96K for GLM-5.2.

API behavior

Model ID: glm-5.3
Thinking is always enabled; disabling it is not supported.
Reasoning effort supports low, high and max; the default is max.
Z.ai recommends Max effort for coding tasks.

Migration note

Applications using thinking.type: "disabled" must change it to enabled before switching model IDs. Use reasoning_effort: "low" when a lighter reasoning budget is needed; unsupported thinking settings cause the request to fail.

Context and availability

Official evaluations use up to a 1M-token context for several long-horizon suites and up to 128K output in selected tests. Z.ai says the open weights will follow two weeks after launch, after safety evaluation and hardening.

Security positioning

The release adds substantial vulnerability-discovery and exploitation capability. Use it only for systems you own or are explicitly authorized to assess, with human review and controlled execution boundaries.

Launch claims and benchmark figures are reported by Z.ai. Private-benchmark results are not independently reproducible from the announcement alone. Read the official GLM-5.3 release ↗

04 / Upcoming change

DeepSeek pricing changes on August 16

Scheduled for August 16, 2026 at 16:00 UTC (9:30 PM IST). Peak rates are twice the announced off-peak rates.

V4 ProReasoning tier

Off-peak

$0.66 in · $1.98 out$0.022 cached input

Peak

$1.32 in · $3.96 out$0.044 cached input

V4 FlashEfficiency tier

Off-peak

$0.22 in · $0.66 out$0.007 cached input

Peak

$0.44 in · $1.32 out$0.014 cached input

05 / Decision guide

Choose for the workload, not just the rate card

A useful shortlist balances unit cost, output volume, reliability and the value of a successful response.

New coding frontierZ.ai

GLM-5.3

Complex coding, long-horizon engineering and authorized cyber-defense work

Uses GLM-5.2’s published rate as a provisional planning proxy until Z.ai publishes a dedicated GLM-5.3 direct API rate.Check provider pricing ↗

Premium capabilityMoonshot AI

Kimi K3

Complex reasoning, long context and high-value agent work

Output-heavy applications should model generated tokens carefully.Check provider pricing ↗

Balanced agentsZ.ai

GLM-5.2

Long coding agents and workloads with reusable context

This comparison uses Z.ai’s published $1.40 / $0.26 / $4.40 pricing tier. Confirm the model identifier before launch.Check provider pricing ↗

Low-cost reasoningDeepSeek

DeepSeek V4 Pro

Reasoning and agent workflows where capability matters

Current rate through August 16, 2026 at 16:00 UTC; scheduled pricing is included.Check provider pricing ↗

Coding specialistMoonshot AI

Kimi K2.7 Code

Repository work, code generation and multi-step coding agents

Coding-focused pricing is attractive when generated output remains controlled.Check provider pricing ↗

Lowest current costDeepSeek

DeepSeek V4 Flash 0731

High-volume assistants, classification and cost-sensitive features

The 0731 release uses the V4 Flash endpoint rate in this snapshot.Check provider pricing ↗

06 / Cost controls

What changes the real bill

↺

Cache stable prefixes

Keep system instructions and reusable context at the beginning. Track actual cache-hit tokens instead of assuming every repeated prompt qualifies.

↧

Control output length

Output is the most expensive token class for most models here. Use clear completion criteria and sensible maximum-token limits.

⑂

Measure agent loops

A single user action may trigger planning, tool calls, retries and verification. Price the complete run, not only the final response.

◌

Route by task

Use an efficient model for routine work and reserve premium models for requests where additional capability creates measurable value.

△

Include failure cost

The cheapest call is not cheapest when it requires repeated attempts. Compare successful-task cost using production evaluation data.

◎

Set usage guardrails

Add per-user quotas, spend alerts and graceful fallbacks. Recalculate after prompt, model or traffic changes.

07 / Method

How the estimate is calculated

monthly requests = requests/day × active days
adjusted tokens = tokens/request × monthly requests × (1 + overhead)
input cost = uncached input × input rate + cached input × cache rate
monthly estimate = (input cost + output cost) × (1 + gateway markup)