Now Live

GPT-5.4 Is Live
in Amplifier

Full ecosystem support — every provider, every matrix, every doc.

Active

March 6, 2026

The Model

What GPT-5.4 Brings

OpenAI's most capable model yet. One model that replaces the entire GPT-5.3-Codex line.
Amplifier supports the core capabilities below, including the full 1.05M context window with reliability streaming. Computer use and tool search are future integration targets.

🧠

Coding + Reasoning, Unified

Merges GPT-5.3-Codex + mainline into one model
5 reasoning levels: none → low → medium → high → xhigh
1M context opt-in (272K standard)

🖥️

Native Computer Use

Reads screenshots, clicks buttons, navigates software
Tool search — lightweight tool list with on-demand definitions
47% token reduction via tool search

💰

Pricing

$2.50 / $15 per 1M tokens (input/output)
43% more per input token vs GPT-5.3
More token-efficient overall — net cost comparable

For Amplifier Users

What You Unlock

Zero configuration changes. Pull latest and everything below is already working.

GPT-5.4 as the default across all OpenAI providers — direct, Azure, and ready for Copilot
Intelligent routing — the routing matrix automatically assigns GPT-5.4 to the right roles with the right reasoning level
Two strategies, one toggle — switch between cost-efficient (balanced) and maximum quality (quality) matrices
Correct context budgets — 272K reported accurately, no silent truncation
xhigh reasoning — the most powerful reasoning level, for critique and architecture tasks
Reasoning token preservation — multi-turn sessions keep full reasoning context across turns

Routing Intelligence

Smart Model Routing

Pick a strategy. The system handles model selection, reasoning levels, and fallback order.

balanced

cost-efficient · production-proven

Role	Model	Reasoning
General chat	gpt-5.4	none
Coding	gpt-5.4	none
Security audit	gpt-5.4	high
Architecture	gpt-5.4-pro	high
Critique	gpt-5.4	xhigh
Research	gpt-5.4-pro	high
Creative	gpt-5.4	none
Fast tasks	gpt-5-mini	—

quality

maximum capability

Role	Model	Reasoning
General chat	gpt-5.4	low
Coding	gpt-5.4	medium
Security audit	gpt-5.4	high
Architecture	gpt-5.4-pro	high
Critique	gpt-5.4-pro	xhigh
Research	gpt-5.4-pro	high
Creative	gpt-5.4-pro	low
Fast tasks	gpt-5-mini	—

Reasoning Levels

The Reasoning Spectrum

GPT-5.4 supports five reasoning levels. The right level is chosen automatically per task.

none Fast, cheap
General tasks

low Light thinking
Quality uplift

medium Balanced
Code generation

high Deep analysis
Security, research

xhigh Maximum depth
Critique, architecture

Key insight: GPT-5.4 defaults to none — you must opt in to reasoning. This is the opposite of GPT-5.2's behavior. Amplifier's routing matrices handle this automatically, so you always get the right reasoning level for the task.

Provider Routing

Multi-Provider Intelligence

The routing matrix tries providers in priority order. GPT-5.4 slots in with correct capabilities.

Agent Request

Role: “coding” · Matrix: balanced

↓

1

Anthropic

claude-sonnet-4-6

if installed

2

OpenAI
gpt-5.4
✦ New

3

Google

gemini-3.1-pro

if installed

4

Ollama

local model

fallback

272K context window reported accurately (not the old 400K)
Reasoning effort per role — none for coding in balanced, medium in quality
Forward-compatible detection — new gpt-5.* models auto-inherit capabilities

Quality

What We Fixed Along the Way

Silent reasoning token loss — discovered during real-device testing. 70 orphaned reasoning references across 4 GPT-5.4 sub-sessions.

The Problem

The provider checked “does this model reason by default?”

GPT-5.4 answer: No — it defaults to none.

But the right question is “CAN this model reason?”

GPT-5.4 answer: Yes — and when instructed, it produces reasoning tokens that need preservation.

The Fix

One conceptual change: gate on capability, not behavior.

Check supports_reasoning (can it reason?) instead of default_reasoning_effort (does it reason by default?).

Matches the Anthropic provider's proven pattern. Shipped as hotfix PR #28.

This is why we test on real devices before shipping.

Architecture

Forward-Compatible by Design

The new ModelCapabilities architecture means the next model drop is easier.

🔮

Auto-inherit capabilities — unknown models in the gpt-5.* family automatically get the latest settings. No emergency patches when new models drop.

📐

Per-model context windows — each model reports its actual limits. No more one-size-fits-all numbers causing silent truncation.

🔗

Unified pattern — Anthropic and OpenAI providers now use the same capability detection approach. Consistent behavior across providers.

🚀

GPT-5.5, GPT-6, whatever comes next — the system is ready. New models inherit, specific overrides are easy.

Impact

By the Numbers

0

Repos Updated

0

PRs Merged

0

Tests Passing

0

Failures

0

Stale Model Strings

0

Silent Bug Caught

Get Started

You're Already on GPT-5.4

# Pull latest — GPT-5.4 is the default amplifier # Switch routing strategies amplifier routing show balanced # cost-efficient amplifier routing show quality # maximum capability # Check what model you’re using amplifier routing show # see the full matrix

provider-openai #27 + #28 · routing-matrix #1 · provider-azure #14 · provider-copilot #23 · foundation #118 · amplifier #230

GPT-5.4 Is Livein Amplifier