48% | Will model context window pricing (cost per million tokens for 1M+ context) decline more than 60% YoY in 2026?

Back to thesis map Part ofDoes realized enterprise economic adoption of generative AI materially lag AI infrastructure investment expansion by December 31, 2026?

Current eventUnit economics & pricing

Will model context window pricing (cost per million tokens for 1M+ context) decline more than 60% YoY in 2026?

LiveResolves Dec 31, 2026

Current view

48%

Current 4 model consensus

Movement

Recent consensus movement

Model disagreement

10pt

Range across the model set

Evidence

Aligned

Models broadly agree on the evidence

Thesis map

What is carrying or resisting the view?

Start with the map. Select a mechanism or condition to see what supports the thesis, what resists it, and where disagreement deserves review.

What changed

No decision-relevant findings are available yet.

Review record

Weekly freezes, daily history, and the map version preserved with the record.

Why the view is fragile

10pt model spread around 48% consensus. Open the full disagreement instrument only when you need the model-level reasoning.

Model disagreement

Model spread

10pt

45% to 55%

Consensus

48%

4 models

Evidence state

Aligned

Models broadly agree

Most divergent

Grok

+7pt from consensus

Disagreement synthesis

The models share a central case but assign different weight to the evidence.

GPT-4o anchors the lower end at 45%, while Grok anchors the upper end at 55%. Consensus sits at 48%, leaving a 10 point range for review.

Primary review question

Whether hyperscale providers maintain 2025–26 gross-margin targets above 70% or accelerate price cuts to capture share.

Shared probability axis

Range 45% to 55%

Consensus 48%

GPT-4o 45

Gemini 45

Claude 50

Grok 55

0255075100

ModelForecastVs consensusDivergenceReasoning signal

Model reasoning

Select a model to inspect

Reasoning

Current 1M+ context pricing sits at $0.60–$3.00 per million tokens (Gemini 1.5 Pro 1M, Claude 3 200K, GPT-4o 128K), while the 2023–2025 compound annual decline has averaged ~45% driven by HBM3e supply ramp and speculative decoding optimizations. 2026 supply schedules show TSMC 2 nm risk production in H2-2025 plus Samsung HBM4 sampling, projecting another 50–70% reduction in inference FLOPs/$ that historically translates to 55–65% price cuts when passed through to list rates.

Key uncertainty

Whether hyperscale providers maintain 2025–26 gross-margin targets above 70% or accelerate price cuts to capture share.

Deep dive

Full decomposition, component audit, trajectory, and detailed analytical views.

Deep dive

Detailed evidence, audit, and history views.

Open only what you need

Belief trajectoryConsensus 48% · 10pt spread

Belief trajectory0 updates

Belief trajectory

Consensus—

Spread—

Confidence—

Δ Consensus—

Loading belief trajectory…

Resolution & provenance

Resolution sources, record policy, cadence, and permanent record metadata.

Resolution and provenance

How the thesis resolves and how its record is maintained.

Permanent record

Resolution sources

Frontier provider pricing pages 2026

Record ID

CRENE-AIER-C087-20261231

Record policy

Append only

Freeze cadence

Thursdays

Generated

Jun 29, 2026

Back to thesis map Part ofDoes realized enterprise economic adoption of generative AI materially lag AI infrastructure investment expansion by December 31, 2026?

Current eventUnit economics & pricing

Will model context window pricing (cost per million tokens for 1M+ context) decline more than 60% YoY in 2026?

LiveResolves Dec 31, 2026

Current view

48%

Current 4 model consensus

Movement

Recent consensus movement

Model disagreement

10pt

Range across the model set

Evidence

Aligned

Models broadly agree on the evidence

Thesis map

What is carrying or resisting the view?

Start with the map. Select a mechanism or condition to see what supports the thesis, what resists it, and where disagreement deserves review.

What changed

No decision-relevant findings are available yet.

Review record

Weekly freezes, daily history, and the map version preserved with the record.

Why the view is fragile

10pt model spread around 48% consensus. Open the full disagreement instrument only when you need the model-level reasoning.

Model disagreement

Model spread

10pt

45% to 55%

Consensus

48%

4 models

Evidence state

Aligned

Models broadly agree

Most divergent

Grok

+7pt from consensus

Disagreement synthesis

The models share a central case but assign different weight to the evidence.

GPT-4o anchors the lower end at 45%, while Grok anchors the upper end at 55%. Consensus sits at 48%, leaving a 10 point range for review.

Primary review question

Whether hyperscale providers maintain 2025–26 gross-margin targets above 70% or accelerate price cuts to capture share.

Shared probability axis

Range 45% to 55%

Consensus 48%

GPT-4o 45

Gemini 45

Claude 50

Grok 55

0255075100

ModelForecastVs consensusDivergenceReasoning signal

Model reasoning

Select a model to inspect

Reasoning

Key uncertainty

Whether hyperscale providers maintain 2025–26 gross-margin targets above 70% or accelerate price cuts to capture share.

Deep dive

Full decomposition, component audit, trajectory, and detailed analytical views.

Deep dive

Detailed evidence, audit, and history views.

Open only what you need

Belief trajectoryConsensus 48% · 10pt spread

Belief trajectory0 updates

Belief trajectory

Consensus—

Spread—

Confidence—

Δ Consensus—

Loading belief trajectory…

Resolution & provenance

Resolution sources, record policy, cadence, and permanent record metadata.

Resolution and provenance

How the thesis resolves and how its record is maintained.

Permanent record

Resolution sources

Frontier provider pricing pages 2026

Record ID

CRENE-AIER-C087-20261231

Record policy

Append only

Freeze cadence

Thursdays

Generated

Jun 29, 2026