Consolidated release bundling today's community fixes.

Added

• LangChain / LangGraph coverage — usage is now tracked from modern chat-model usage_metadata and Anthropic response_metadata (not just legacy token_usage), so chat models and LangGraph runs are counted instead of silently recording $0.00. Adds tool-cost tracking, a context-manager lifecycle that fires on_hard_limit, and callback forwarding. Thanks @veeradyani222! (#26)
• InvalidCost exception, exported from the package.
• Contributor docs — issue/PR templates and CONTRIBUTING.md. Thanks @sundaram2021! (#25)

Changed

• Drop-in API isolation — init()/teardown() session state is now scoped per thread and per async task via contextvars; concurrent requests no longer overwrite each other's sessions. Thanks @sundaram2021! (#20)

Fixed

• Streaming cost on early exit — recorded when the caller breaks early or an async stream ends incompletely; OpenAI auto-injects include_usage. Thanks @Kunal1522! (#15)
• Invalid cost values — negative/NaN/infinite costs are rejected before corrupting the ledger. Thanks @drewster99! (#21)

CI now also validates Go, TypeScript, and the website, and gates publishing on the full matrix (#24).

Full changelog: https://github.com/AgentBudget/agentbudget/blob/main/CHANGELOG.md

What's new in v0.3.0

Features

• Streaming cost tracking — stream=True calls are now fully tracked for OpenAI (sync + async) and Anthropic (sync + async).
  Chunks pass through unchanged; cost is recorded after the stream is exhausted. Thanks @KTS-o7!

  Note: OpenAI requires stream_options={"include_usage": True} — without it cost is silently $0.

• agentbudget.wrap_client(client, session) — explicit per-client tracking as an alternative to global patching. Only the
  wrapped instance is tracked.
• finalization_reserve param on AgentBudget — reserves a fraction of budget for the final response step, preventing
  agents from being cut off mid-task.
• session.would_exceed(cost) — pre-flight budget check without recording anything.

Bug fixes

• OpenRouter support — model names like "openai/gpt-4o" now resolve correctly. Previously these silently tracked as $0.00.
• Thread safety in CircuitBreaker soft limit check (merged from @MistakenPirate)
• Off-by-one in loop detection threshold (merged from @MistakenPirate)
• Exception matching uses issubclass() instead of string comparison (merged from @MistakenPirate)
• Negative price validation in register_model() (merged from @MistakenPirate)

CI

• Tests now run on every push/PR across Python 3.9–3.13

Full changelog: v0.2.3...v0.3.0

What's New

Custom Model Pricing — Don't Wait for a Release

New model just launched? Register pricing at runtime:

import agentbudget

agentbudget.register_model(
    "gpt-5",
    input_price_per_million=5.00,
    output_price_per_million=20.00,
)

# Or batch register:
agentbudget.register_models({
    "gpt-5": (5.00, 20.00),
    "gpt-5-mini": (0.50, 2.00),
})

Custom pricing takes priority over built-in pricing, so you can also override existing models.

Fuzzy Matching for Dated Model Variants

Dated model names like gpt-4o-2025-06-15 now automatically match their base model (gpt-4o) pricing. No more None costs for new model
versions.

Resolution order: Custom pricing → Built-in table → Fuzzy match.

Updated Docs & README

• New "Custom Model Pricing" section in docs
• register_model and register_models added to API reference
• Updated supported models table (40+ models)

Install / Upgrade

  pip install --upgrade agentbudget

Full Changelog: v0.2.2...v0.2.3

What's New

Drop-in Auto-Instrumentation

Two lines to enforce a budget on any AI agent session — zero code changes needed:

import agentbudget
agentbudget.init("$5.00")

Your existing OpenAI/Anthropic code works unchanged

Updated Model Pricing (Feb 2026)

Added pricing for 15+ new models:

• OpenAI: GPT-4.1, GPT-4.1-mini, GPT-4.1-nano, o3, o3-pro, o4-mini
• Anthropic: Claude Opus 4.6, Opus 4.5, Sonnet 4.5, Sonnet 4, Haiku 4.5
• Google: Gemini 2.5 Pro, Gemini 2.5 Flash, Gemini 2.5 Flash Lite
• Mistral: Updated pricing for mistral-large and mistral-small

CI/CD

• Automated PyPI publishing via GitHub Actions with trusted publishing
• Test matrix across Python 3.9, 3.10, 3.12, 3.13

Project

• New domain: agentbudget.dev
• Added project URLs to PyPI metadata (docs, repo, changelog)

Install / Upgrade

pip install --upgrade agentbudget

Full Changelog: v0.1.0...v0.2.1

What's New

Drop-in Auto-Instrumentation

Two lines to enforce a budget on any AI agent session — zero code changes needed:

import agentbudget
agentbudget.init("$5.00")

Your existing OpenAI/Anthropic code works unchanged

Updated Model Pricing (Feb 2026)

Added pricing for 15+ new models:

• OpenAI: GPT-4.1, GPT-4.1-mini, GPT-4.1-nano, o3, o3-pro, o4-mini
• Anthropic: Claude Opus 4.6, Opus 4.5, Sonnet 4.5, Sonnet 4, Haiku 4.5
• Google: Gemini 2.5 Pro, Gemini 2.5 Flash, Gemini 2.5 Flash Lite
• Mistral: Updated pricing for mistral-large and mistral-small

CI/CD

• Automated PyPI publishing via GitHub Actions with trusted publishing
• Test matrix across Python 3.9, 3.10, 3.12, 3.13

Project

• New domain: agentbudget.dev
• Added project URLs to PyPI metadata (docs, repo, changelog)

Install / Upgrade

pip install --upgrade agentbudget

Full Changelog: v0.1.0...v0.2.1

What's New

Drop-in Auto-Instrumentation

Two lines to track every OpenAI and Anthropic call automatically — zero changes to your existing code:

import agentbudget
agentbudget.init("$5.00")

Updated Model Pricing (February 2026)

Added 20+ new models including GPT-4.1, GPT-4.1-mini, GPT-4.1-nano, o3, o4-mini, Claude Opus 4.6, Claude Sonnet 4.5, Claude Haiku 4.5,
Gemini 2.5 Pro/Flash. Updated Mistral pricing.

Website & Docs

• Live at agentbudget.org
• Full documentation at agentbudget.org/docs

Other Improvements

• Async session support (budget.async_session())
• Nested budgets with automatic cost rollup
• Webhook support for budget events
• LangChain and CrewAI integrations
• Loop detection and circuit breaker
• Event callbacks (on_soft_limit, on_hard_limit, on_loop_detected)

Install / Upgrade

pip install --upgrade agentbudget

Full Changelog: v0.1.0...v0.2.0

AgentBudget v0.1.0

Real-time cost enforcement for AI agent sessions.

Features

• Per-session dollar-denominated budget enforcement
• Auto-costing for OpenAI, Anthropic, Google Gemini, Mistral, and Cohere models
• Manual cost tracking for tool calls and external APIs
• Circuit breaker with hard limit, soft limit warnings, and loop detection
• Async support
• Nested budgets (parent/child sessions)
• Webhook event streaming
• LangChain callback and CrewAI middleware integrations
• JSON cost reports

Install

pip install agentbudget

Quick Start

from agentbudget import AgentBudget

budget = AgentBudget(max_spend="$5.00")
with budget.session() as session:
response = session.wrap(openai_call(...))
session.track(tool_result, cost=0.01)
print(session.report())