grok-4.5 and grok-4.20-0309-reasoning: internal reasoning
On this page
Advanced reasoning models in 2026
xAI’s high-end reasoning lineup currently relies on two models: grok-4.5, the flagship model released on July 16, 2026, with configurable reasoning, and grok-4.20-0309-reasoning, the variant explicitly dedicated to reasoning. Unlike the previous grok-3-mini generation, their thought process is not exposed in plain text by default: you receive the final answer, while the reasoning remains internal.
This does not mean reasoning is absent. On the contrary, these models employ reasoning capabilities significantly superior to the previous generation. The difference lies in visibility: the reasoning exists, but it is not delivered as readable text.
Why internal reasoning?
Not exposing raw reasoning serves several technical and strategic objectives:
- Intellectual property protection: the internal reasoning process constitutes an innovation that xAI protects
- Extraction risk reduction: prevent reverse engineering of the model’s thought mechanisms
- Optimization: a compact format (possibly encrypted) allows efficient reuse between requests
From a developer perspective, you cannot directly inspect the reasoning. However, the API has historically allowed retrieving reasoning in encrypted form to reinject it into subsequent requests.
Retrieving encrypted reasoning
The mechanism goes through the Responses API with the include parameter:
from openai import OpenAI
client = OpenAI(
api_key="your-api-key",
base_url="https://api.x.ai/v1"
)
response = client.responses.create(
model="grok-4.20-0309-reasoning",
input="Prouvez que la somme des angles d'un triangle est 180 degrés.",
include=["reasoning.encrypted_content"]
)The response then contains a field with the encrypted reasoning content: an encoded blob that you cannot read, but that you can store and reuse in a later request to preserve reasoning continuity.
Check the official models page (docs.x.ai) to confirm that the model you are targeting supports this field: API option availability may vary from one model to another.
Models and their characteristics
grok-4.5
xAI’s flagship model since July 16, 2026, deployed on iOS, Android, the web, and X since July 22. 500k token context, configurable reasoning: you choose the reasoning effort based on the task. It is also the default model for Grok Build. Tiered pricing: $2 input and $6 output per million tokens under 200k context, $4/$12 beyond.
grok-4.20-0309-reasoning
The reasoning-dedicated variant of the grok-4.20-0309 family, still available in the catalog. 1 million token context, significantly lower rates than grok-4.5: $1.25/$2.50 per million tokens under 200k context, $2.50/$5 beyond. This is the economical choice for long reasoning chains or very large contexts.
grok-4.20-multi-agent-0309
The third variant of the 4.20 family: oriented toward multi-agent orchestration, with the same 1M context and same rates as the reasoning variant. Useful when your architecture makes several agents collaborate rather than a single model reasoning at length.
Note: grok-4 without suffix and grok-4-fast variants belong to previous generations and no longer appear in the current catalog.
grok-4.5 vs grok-4.20-0309-reasoning
| Feature | grok-4.5 | grok-4.20-0309-reasoning |
|---|---|---|
| Role | Flagship model, configurable reasoning | Reasoning-dedicated variant |
| Context | 500k tokens | 1M tokens |
| Input/Output Price (< 200k) | $2 / $6 per MTok | $1.25 / $2.50 per MTok |
| Input/Output Price (≥ 200k) | $4 / $12 per MTok | $2.50 / $5 per MTok |
| Reasoning visible in plain text | Not guaranteed — check docs | Not guaranteed — check docs |
| Typical use case | Complex problems, maximum quality | Volume, long contexts, controlled budget |
Pricing as of August 5, 2026 — prices change regularly: before any budget calculation, check the current rate card on the official xAI models and pricing page.
Key takeaways
- In 2026, xAI advanced reasoning relies on grok-4.5 (configurable, 500k) and grok-4.20-0309-reasoning (dedicated, 1M)
- These models’ reasoning is internal: you receive the final answer, not the chain of thought in plain text
- Encrypted reasoning was historically retrieved via
include=["reasoning.encrypted_content"]to be reused between requests (lesson 7) — check official docs to confirm the targeted model still supports it - grok-4.20-0309-reasoning costs less than half of grok-4.5 for output under 200k context ($2.50 vs $6 per MTok)
- grok-4 without suffix and grok-4-fast are previous generations: do not use them in new projects