← Agent Frameworks 🕐 20 min read
Agent Frameworks

AI Model Price Deflation — Historical Trajectory and Commitment Timing Strategy (2026)

Executive Summary · March 2026

AI inference pricing has undergone a structural collapse with no precedent in enterprise software procurement.

AI inference pricing has undergone a structural collapse with no precedent in enterprise software procurement. GPT-4’s input token cost fell from $30/M at launch (March 2023) to $2.50/M by May 2024 — a 92% drop in 14 months. The trend has continued: by June 2026, equivalent frontier capability is available at $1–3/M input tokens across multiple providers, and budget-tier capable models have dropped to $0.01–0.15/M. The overall output-token price for frontier models has fallen approximately 94% since 2023.

This is not temporary competitive discounting. It is driven by compounding structural forces — hardware efficiency (H100 → B200), algorithmic improvements (FlashAttention, speculative decoding, MoE architectures), quantization, and open-source competition anchoring the floor. Applying Wright’s Law to cumulative token production, inference cost tracks a power-law decay, and the slope is steeper than classical manufacturing curves: approximately 10× cost reduction per year for equivalent capability.

For enterprise FinOps teams, this creates a procurement paradox: committing to capacity locks in today’s price in a market where prices fall materially every quarter. This note quantifies the historical trajectory, explains the structural drivers, models the forward curve, and provides a commitment-timing framework.

1. Historical Price Data: The Collapse in Detail

1.1 OpenAI Model Pricing Timeline

Date Model Input ($/M) Output ($/M) Notes
Mar 2023 GPT-4 (8K) $30.00 $60.00 Launch pricing
Nov 2023 GPT-4 Turbo $10.00 $30.00 67% input reduction vs GPT-4 8K
May 2024 GPT-4o $5.00 $15.00 Launch; 50% cut vs GPT-4 Turbo
Jul 2024 GPT-4o (updated) $2.50 $10.00 Second cut; 50% again in 8 weeks
Jul 2024 GPT-4o mini $0.15 $0.60 Replaced GPT-3.5-Turbo
Jan 2025 GPT-4.5 $75.00 $150.00 Flagship; short-lived at this price
Mid-2025 GPT-4.5 (revised) $5.00 $15.00 Rapid repricing post-launch
2026 GPT-5.4 (illustrative) $2.50 $10.00 ~12× reduction from GPT-4 launch

Key metric: GPT-4 input cost: $30/M (Mar 2023) → $2.50/M (May 2024). 92% drop in 14 months.

Cumulative GPT-4 class deflation (Mar 2023 → Jun 2026): Input tokens approximately 12× cheaper. Output tokens approximately 6–8× cheaper (output deflation lagged slightly).

1.2 Anthropic Claude Pricing Timeline

Date Model Input ($/M) Output ($/M) Notes
Mar 2023 Claude 1 ~$11.02 ~$32.68 Initial API access, limited availability
Jul 2023 Claude 2 $11.02 $32.68 General availability
Nov 2023 Claude 2.1 $8.00 $24.00 Incremental reduction
Mar 2024 Claude 3 Opus $15.00 $75.00 Flagship launch; premium positioning
Mar 2024 Claude 3 Sonnet $3.00 $15.00 Mid-tier; strong value positioning
Mar 2024 Claude 3 Haiku $0.25 $1.25 Fastest, cheapest tier
Jun 2024 Claude 3.5 Sonnet $3.00 $15.00 Capability step-up at same Sonnet price
Oct 2024 Claude 3.5 Haiku $0.80 $4.00 Haiku tier capability upgrade
Feb 2025 Claude 3.7 Sonnet $3.00 $15.00 Extended thinking; same Sonnet price
Q3 2025 Claude 4.1 Opus $15.00 $75.00 Legacy pricing on prior flagship
Q4 2025 Claude 4.7 Opus $15.00 $75.00 Fast Mode: $30/$150
May 2026 Claude Opus 4.8 $5.00 $25.00 67% cut from Opus 4.7 standard
May 2026 Claude Opus 4.8 (Fast) $10.00 $50.00 Down from $30/$150 on 4.7
May 2026 Claude Sonnet 4.6 $3.00 $15.00 Unchanged price; 4 generations held
May 2026 Claude Haiku 4.5 $1.00 $5.00 4× more expensive than Haiku 3

Key observation: The Sonnet line has held $3/$15 across four generations (Sonnet 3.7, 4.1, 4.5, 4.6) while capability increased substantially. This is deflation embedded in capability, not price — the price is flat but the model is significantly more capable. Adjusted for capability, the Sonnet line has deflated materially.

Opus pricing: $15 (Claude 3 Opus, Mar 2024) → $5 (Claude Opus 4.8, May 2026). 67% reduction in 26 months.

1.3 Budget Model Pricing: The Floor Descends

Date Model Input ($/M) Output ($/M)
Jul 2024 GPT-4o mini $0.15 $0.60
Dec 2024 Google Gemini Flash 1.5 $0.075 $0.30
Jan 2025 DeepSeek V3 (hosted) $0.27 $1.10
Mar 2025 Gemini 2.0 Flash $0.10 $0.40
2025 DeepSeek V3 (bulk) $0.07 $0.28
2026 DeepSeek V4 Flash $0.01 $0.03
Jun 2026 Median budget frontier $0.10–0.15 $0.40–0.60

The budget-tier floor has effectively reached a cent-per-million-token threshold for capable models. For many enterprise workloads, the cost bottleneck has shifted from model pricing to orchestration overhead and latency management.

1.4 Deflation Rate Summary

Period Metric Deflation
Mar 2023 – May 2024 (14 mo) GPT-4 class input $/M −92%
Mar 2023 – Jun 2026 (39 mo) GPT-4 class input $/M ~−92% (additional 50% from GPT-4o further cuts)
Mar 2024 – Jun 2026 (26 mo) Claude Opus $/M input −67%
2023 – 2026 Average frontier output token price ~−94%
2024 – 2026 Budget-tier floor ~−90%+
Annualized (2023–2026) Frontier input price ~−60% to −70%/year

2. What Drives the Deflation

Deflation is not a market anomaly or temporary promotional pricing. It results from five compounding structural forces, each of which continues to operate.

2.1 Inference Algorithm Efficiency

FlashAttention and attention kernel optimization: FlashAttention (2022) and FlashAttention-2 (2023) reduced memory I/O for the attention mechanism by 5–10×, enabling longer context at lower cost. These gains compound with hardware improvements because attention was the primary memory-bandwidth bottleneck.

Speculative decoding: A small draft model proposes multiple token candidates that the large target model then verifies in parallel. This increases throughput 2–3× with no quality loss. Speculative decoding is now standard practice at major inference providers.

Continuous batching: Replaced static batching (which padded sequences to the longest batch member), recovering 50–80% of wasted GPU cycles on typical workloads.

PagedAttention / KV cache management: vLLM’s PagedAttention (2023) eliminated KV cache fragmentation and memory waste. Combined with quantized KV caches, this reduces memory overhead by 2–4× and allows more concurrent requests per GPU.

Combined algorithmic effect: Software optimization alone has reduced the compute cost of serving a given model by 2–3× from 2023 to 2026, independent of hardware.

2.2 Quantization

Running model weights and activations at reduced numerical precision cuts memory requirements and increases throughput with minimal accuracy loss:

Precision Throughput gain vs FP32 Quality loss
FP16 ~2× Negligible
FP8 (H100) ~4× vs FP32, 1.3–2× vs FP16 <2% on instruction-tuned models
FP4 (B200) ~6–8× vs FP32 2–4% on instruction-tuned models
INT4 (AWQ/GPTQ) ~4–6× 1–3% on standard benchmarks

B200 hardware running FP4 via TensorRT-LLM adds a 1.5–2× additional gain over FP8 on H100. As providers migrate fleets to Blackwell, this gain flows directly into margin and pricing.

2.3 Mixture-of-Experts (MoE) Architectures

MoE models activate only a fraction of total parameters per token. DeepSeek V3 (671B total, 37B active per token) delivers performance competitive with dense 70B models at roughly 1/18th the active compute per token. GPT-4o and Gemini 1.5 are widely believed to use MoE architectures. This architectural shift is the single largest driver of frontier model cost reduction in 2024–2025 — it changes the fundamental economics of serving frontier-class capability.

Effect on inference cost: MoE brought an estimated 18× reduction in per-token compute requirements for frontier-class output from 2023 to 2026, per academic analysis of architectural transitions (arxiv 2603.28576).

2.4 Hardware Cost Reduction

H100 cloud pricing trajectory:

  • 2023: $7–8/GPU-hr (spot)
  • 2024: $2.50–4.00/GPU-hr
  • 2025: $1.49–3.90/GPU-hr (AWS cut prices 44% in June 2025)
  • 2026: Blackwell GB200 NVL72 racks available; 30× inference throughput improvement vs H100 per NVIDIA specification

Each hardware generation compresses cost independently of software. B200 vs H100 represents the largest per-generation efficiency leap in Nvidia’s history for inference workloads.

2.5 Competition as Structural Floor-Setter

The competitive dynamic is asymmetric: A single low-cost entrant (DeepSeek V3, January 2025) forced repricing across the entire industry within weeks. Google, Anthropic, OpenAI, and Microsoft all made pricing adjustments in Q1 2025 in direct response. The open-weight ecosystem (Llama 4, Qwen 3.5, DeepSeek V4) anchors the floor — any enterprise can self-host comparable capability, so hosted providers cannot price above what self-hosting would cost at scale.

The self-hosting anchor: For a 1,000-GPU-hr/day enterprise workload, self-hosting a Qwen 3.5 or Llama 4 class model on Blackwell hardware at current spot prices costs approximately $0.10–0.20/M input tokens all-in (compute + engineering + inference serving overhead). This is the structural ceiling for hosted budget-tier pricing.

3. Wright’s Law Applied to AI Inference

3.1 The Classical Framework

Wright’s Law (1936, Theodore Wright): for every cumulative doubling of units produced, cost falls by a constant percentage — the “learning rate.” Solar photovoltaics demonstrated an 18–20% learning rate (cost falls 18–20% per doubling of cumulative installed capacity). Lithium-ion batteries: ~18%. DRAM: ~32%.

Applied to AI accelerators: ARK Invest analysis corroborated Wright’s Law for GPU compute, finding a 37.5% cost decline for every cumulative doubling in AI compute units produced.

3.2 AI Inference Exceeds Classical Wright’s Law

Standard Wright’s Law captures manufacturing learning but misses algorithmic improvement. AI inference has both:

  1. Hardware Wright’s Law: ~37.5% reduction per cumulative doubling of GPU production
  2. Algorithmic efficiency: Separate curve — software improvements independent of hardware scale
  3. Architecture step-changes: MoE, attention improvements — discrete jumps not captured by smooth curves

The empirical result is that frontier inference cost has fallen approximately 10× per year for equivalent capability. This is approximately 3× faster than Wright’s Law on hardware alone would predict.

Power-law fit (approximate): If $C_0$ is cost at time 0 and $t$ is years elapsed:

$$C(t) = C_0 \cdot (0.1)^t$$

i.e., cost falls to 10% of its current level per year for equivalent-capability inference. This is an empirical fit; the actual rate may compress as hardware roadmaps slow and algorithmic low-hanging fruit is exhausted.

3.3 Forward Curve Projection

Starting from June 2026 frontier pricing ($3/M input for Sonnet 4.6 class; $0.10/M for budget frontier):

Year Frontier equivalent ($/M input) Budget equivalent ($/M input)
2026 (now) $3.00 $0.10
2027 $0.30–1.00 $0.01–0.03
2028 $0.10–0.30 $0.001–0.005
2029 $0.05–0.15 <$0.001

Confidence interval note: The 10×/year rate is empirical through 2026, but several countervailing forces could slow it: H100 → B200 transition completion (a one-time step), benchmark saturation, reasoning-intensive workloads that require more compute per token (o3-class models), and market consolidation. A conservative floor scenario assumes 3–5×/year deflation from 2027 onward.

The floor is not zero. Inference requires compute, which requires energy. Energy cost sets a hard floor. For a B200 GPU at $1/GPU-hr, running at peak throughput of ~10M tokens/hr, the compute floor is $0.0001/M tokens — effectively zero for enterprise pricing purposes, but not for hyperscale self-hosters.

4. Capability Inflation Alongside Price Deflation

4.1 The Compounding Ratio

Price is falling, but capability is rising simultaneously. The relevant enterprise metric is not price per token but capability per dollar.

Benchmark improvement rate: The Epoch AI Capabilities Index shows the best frontier score grew approximately 8 points/year before April 2024, accelerating to ~15 points/year after — roughly a 90% acceleration in improvement rate. SWE-bench Verified (software engineering) went from ~60% (mid-2024) to near 100% (2025). GPQA went from ~50% to 75%+ in 18 months.

Performance per dollar: Conservative estimates put improvement at approximately 30% per year on performance-normalized metrics. Combined with the price deflation rate, the effective capability per dollar is improving 40–60× per year at the frontier.

4.2 What This Means for Enterprise Budgeting

A workload priced at current rates and current capability assumptions will be:

  • Executable at 10× less cost in 12 months with equivalent or superior output quality
  • Or, for the same cost, executable at 10× the volume or with significantly higher-quality reasoning

This creates a strategic asymmetry: committing to a workload definition today locks in assumptions about what $X buys that will be dramatically wrong in 18 months. The correct planning posture is to architect for volume scale-up within a fixed budget, not to lock in volume commitments at current cost.

4.3 Capability Inflation in Reasoning-Intensive Tasks

A critical counterforce: reasoning-heavy tasks (chain-of-thought, extended thinking, agentic loops) use more tokens per task than simple completions. As enterprises move from generation to reasoning workloads, per-task cost may not fall as fast as per-token cost. A multi-step agent loop that runs 50,000 tokens per task is still expensive even at $0.10/M.

Implication for commitment strategy: Model per-task token consumption, not just per-token price. The tokens-per-task metric is partially controlled by prompt engineering and architecture, and is a more stable basis for cost modeling than raw price/token.

5. Commitment Timing: The Risk of Locking In

5.1 Available Commitment Products (June 2026)

Azure OpenAI — Provisioned Throughput Units (PTUs):

  • Monthly commitment: ~15–20% discount vs pay-as-you-go
  • Yearly commitment: 40–65% discount vs pay-as-you-go
  • Break-even: ~300,000 tokens/minute sustained throughput for 8+ hours/day
  • Limitation: PTUs are model-version specific — a commitment to GPT-4o PTUs does not transfer to GPT-5 when it launches

AWS Bedrock:

  • Standard (on-demand), Flex (50% discount with usage commitment), Reserved tiers
  • 1-month minimum commitment on reserved capacity
  • Integrates with AWS Enterprise Discount Program (EDP) — effective discount increases 8–15% when folded into EDP

Anthropic (direct):

  • Batch API: 50% discount, 24-hour turnaround, suitable for async workloads
  • Prompt caching: 90% reduction on cached input tokens — functionally a commitment substitute for workloads with stable system prompts
  • Direct enterprise contracts: negotiated rates, typically volume-tiered

5.2 The Commitment Timing Problem

The mathematical problem: A 12-month PTU commitment at today’s Azure pricing offers 40–65% savings versus on-demand. But if the provider drops the on-demand price 50% in month 3 (as has happened repeatedly), the committed rate becomes more expensive than the new on-demand rate in months 4–12.

Historical evidence of in-period repricing:

  • OpenAI cut GPT-4o price 50% within 8 weeks of launch (May → July 2024)
  • Anthropic cut Claude Opus 4.8 pricing 67% vs prior Opus with May 2026 launch
  • DeepSeek January 2025 release triggered within-quarter repricing across all major providers

The 2024 PTU underutilization finding: Across cost reviews conducted 2024–2025, provisioned throughput commitments ran 30–60% underutilized in the first two quarters after purchase. The combination of overestimated usage and price drops makes 12-month commitments structurally risky in this environment.

5.3 Break-Even Analysis by Commitment Length

Assumptions: Current on-demand price = $3/M input. Expected deflation = 50% in 12 months (conservative; actual has been 60–70%/yr).

Commitment Discount Committed rate On-demand in month 12 Break-even month
1-month ~15% $2.55 $2.55 (same) Month 2
3-month ~25% $2.25 $2.15 (approx) Month 2–3
6-month ~35% $1.95 $1.50 (approx) Month 4
12-month 40–65% $1.05–1.80 $1.50 (approx) Month 6–10

Interpretation: At a 50%/year deflation rate, a 12-month commitment at 50% discount breaks approximately even over the term. At a 70%/year deflation rate (the recent historical pace), a 12-month commitment loses value in the second half. The only commitment that consistently beats on-demand deflation is 1-month or shorter.

Exception — batch workloads: Batch API (Anthropic) and off-peak processing (Azure) offer 50% discounts with no term commitment. For async workloads (document processing, evaluation runs, overnight analysis), batch is structurally superior to reserved capacity at any term length.

Default stance: avoid term commitments on frontier models. The deflation rate exceeds the discount rate for commitments longer than 3 months at current trajectories.

Where commitments make sense:

  1. Predictable baseline at budget-tier models. If your workload runs entirely on GPT-4o mini or Haiku-class models and you have >300K tokens/minute sustained, 3-month commitments are defensible — budget-tier deflation is slower than frontier (floor proximity) and the discount is material.

  2. Batch and async workloads. Use batch APIs (50% discount, no commitment) rather than reserved capacity. Latency tolerance is the constraint; price is not.

  3. Prompt caching as pseudo-commitment. For workloads with large, stable system prompts, prompt caching (90% reduction on cached tokens) is more valuable than any reserved capacity tier and requires no commitment. A 10,000-token system prompt cached across 1M calls saves ~$27,000 at Sonnet 4.6 rates.

  4. EDP/MACC-folded spend. If the enterprise has an existing AWS EDP or Azure MACC, fold AI spend in rather than negotiating a separate AI commitment. The incremental discount (8–15%) comes with no additional AI-specific risk.

  5. New model launch windows. The 90 days after a major model launch (GPT-5, Claude 5) are the worst time to commit. Historical pattern: launch prices drop 30–50% within the first 6 months as the model is optimized for inference and competition responds.

6. Price Floor Analysis: Where Does Inference Commoditize?

6.1 Current Floor Structure

The open-weight ecosystem is the structural floor. DeepSeek V4 Flash at $0.01/M input (hosted) implies a self-hosting cost even lower for hyperscale operators. The hosted price cannot fall below the cost to serve the model on commodity hardware — but commodity hardware itself is deflationary.

Minimum compute cost for a 70B-class model on B200:

  • B200 GPU: ~$1.50/GPU-hr (spot, 2026 estimate)
  • Peak throughput for 70B FP8 on B200: ~8–12M tokens/hr
  • Compute floor: $0.12–0.19/GPU-hr ÷ 10M tokens/hr = $0.012–0.019/M tokens

At this compute floor, current budget-tier pricing ($0.10–0.15/M) still has a 5–10× margin. There is room for additional deflation in the budget tier.

For frontier models (Opus 4.8 class, ~$5/M): These models are not 70B equivalents. They require more compute per token (larger active parameter counts or extended chain-of-thought compute). The floor is higher — estimated $0.50–2.00/M for frontier-class inference on Blackwell, depending on context length and batch efficiency. Current frontier pricing at $3–5/M is approaching margin compression but not yet at floor.

6.2 Reasoning Models — The Cost Exception

Extended-thinking models (o3, Claude 3.7 extended, GPT-4.5 with reasoning) use compute-time search rather than a single forward pass. Per-task costs are higher because more tokens are consumed per answer.

Model class Tokens/task (simple) Tokens/task (complex) Cost/task (complex)
Sonnet 4.6 500–2,000 5,000–20,000 $0.075–$0.30
Claude extended thinking 1,000–5,000 20,000–100,000 $0.30–$1.50
o3-class 2,000–10,000 50,000–500,000 $0.50–$5.00+

Reasoning models inflate per-task cost even as per-token cost falls. This is the key distinction enterprises must track: monitor per-task cost, not per-token price.

6.3 The Commoditization Threshold

Commoditization occurs when the product becomes undifferentiated. AI inference is approaching commoditization for:

  • Routine generation (summarization, classification, extraction) — already commoditized
  • Code generation at the function level — commoditizing in 2026
  • Not yet commoditized: frontier reasoning, multimodal understanding, long-context coherence, agentic reliability

Procurement implication: Treat commoditized workload tiers as cloud compute — buy on spot, no commitments, vendor-agnostic. Treat frontier/reasoning tiers as strategic capability — negotiate on flexibility, not price, because the capability gap between providers still matters.

7. FinOps Implications of a Deflationary Environment

7.1 Why Static AI Budgets Are Wrong

The FinOps Foundation 2025 report found average monthly enterprise AI spend hit $62,964 in 2024, projected to rise to $85,521 in 2025 — a 36% YoY increase. But these projections were built on static-price assumptions. In a deflationary environment:

  • Volume will increase faster than budget implies. If prices fall 60% and the AI budget is flat, the enterprise can run 2.5× the workload volume with the same budget. Most teams will take the volume rather than the savings — which means AI spend stays flat or rises, even as price/token falls.
  • Static budgets underestimate volume growth. A budget built on “10M tokens/day at $3/M = $30/day” is accurate for day one. By month 12, if the enterprise scales to 50M tokens/day at $1/M, actual spend is $50/day — 67% over a static budget based on the original price.
  • The correct budget model is volume-first, not price-first. Estimate the token volume the team will actually consume if price were not a constraint, then multiply by the forward price curve rather than today’s price.

7.2 A Volume-First Budget Model

Step 1: Estimate unconstrained token demand
  - Map each use case to token volume if fully deployed
  - Sum across use cases to get "full deployment volume" (FDV)

Step 2: Apply forward price curve
  - Months 1–3: current price × 1.0
  - Months 4–6: current price × 0.75 (25% deflation)
  - Months 7–12: current price × 0.50 (50% cumulative deflation)

Step 3: Apply adoption ramp
  - Teams do not go from 0 to FDV instantly
  - Typical enterprise adoption ramp: 10% → 40% → 70% of FDV over 12 months

Step 4: Multiply FDV × adoption ramp × forward price
  - This is the expected AI API spend budget

Step 5: Add 20% buffer for volume surprise
  - Price surprises (drops) reduce spend; volume surprises (usage explosions) increase it
  - Volume surprises have historically exceeded volume forecasts by 2–3× in AI deployments

Example (1,000-seat enterprise, mix of coding + summarization + search):

  • FDV: 500M tokens/day = 15B tokens/month
  • Month 1–3 at $2/M blended: 15B × $2 × 10% adoption = $30,000/month
  • Month 4–6 at $1.50/M, 40% adoption: $90,000/month
  • Month 7–12 at $1.00/M, 70% adoption: $105,000/month
  • 12-month total: ~$750,000 — versus a static-price model yielding $450,000 (volume underestimated) or $1.2M (price not deflated)

7.3 Governance Posture in a Deflationary Market

Do not lock in vendor commitments >3 months. The discount does not compensate for price-drop exposure beyond that window based on historical patterns.

Negotiate for price-match clauses in enterprise agreements. If the provider drops the on-demand price, the enterprise agreement rate should automatically follow. This is achievable for large-volume customers and eliminates the primary risk of committing.

Separate infrastructure from inference in cost tracking. Azure OpenAI PTUs conflate capacity management and model access. Track separately: the cost of reserved GPU hours versus the cost of model API calls. This enables clearer make-vs-buy analysis when open-weight self-hosting becomes cost-competitive.

Set per-workload cost-per-task targets, not cost-per-token targets. As prices fall, teams will adopt more expensive models for the same task if the per-task cost stays within budget. A classification task that runs on Haiku at $0.001/task should not migrate to Sonnet just because Sonnet got cheaper — unless the quality improvement is material.

Treat prompt caching as a first-order cost lever. At 90% reduction on cached input tokens, caching a 20,000-token system prompt across 100K daily calls saves ~$54/day at Sonnet 4.6 rates — $19,700/year per high-volume deployment. This is often more valuable than any reserved capacity discount.

8. Vendor Price Change Risk: Contract and Commitment Implications

8.1 Historical Repricing Events and Their Speed

Date Event Speed of impact
Jul 2024 OpenAI cuts GPT-4o 50% 8 weeks after launch
Jan 2025 DeepSeek V3 forces industry repricing Within 2–4 weeks of release
Jun 2025 AWS cuts GPU cloud prices 44% Effective immediately
May 2026 Anthropic cuts Opus 4.8 67% vs Opus 4.7 At model launch

The pattern is consistent: major repricing events arrive without warning and take effect immediately. There is no advance notice period that would allow a committed customer to adjust.

8.2 Structural Risk in Enterprise Agreements

The PTU/reserved capacity trap: Azure PTU pricing is set at commitment time. If Azure reprices GPT-4o by 40% in month 4 of a 12-month PTU commitment, the committed customer continues paying the original rate. There is typically no repricing provision in PTU agreements.

The model-version risk: A 12-month commitment to GPT-4o PTUs expires worthless when GPT-5 launches and becomes the preferred model. Enterprises then face a choice between paying for deprecated model capacity or exiting the commitment at penalty.

The underutilization trap: 30–60% PTU underutilization observed in enterprise deployments implies that organizations systematically overestimate demand at commitment time. The combination of overestimated demand and unexpected price drops means many PTU commitments have negative economic value versus pay-as-you-go.

8.3 Contract Provisions to Negotiate

For any enterprise AI agreement with committed capacity, negotiate the following provisions:

1. Price-match guarantee: If the provider’s standard on-demand price for the committed model drops below the committed rate, the enterprise rate automatically adjusts downward. This is the single most important clause and is achievable for customers committing >$500K/year.

2. Model substitution right: If a newer model in the same family launches, committed capacity credits transfer at parity. Without this, a 12-month GPT-4o commitment becomes a liability when GPT-5 launches.

3. Quarterly volume true-up: Committed volumes should have a quarterly true-up option (up or down by 25%) rather than a fixed annual commitment. Downward flexibility is the critical provision given the overestimation pattern.

4. Competitive matching clause: If a direct competitor offers equivalent capability at a lower price, the enterprise has a 30-day window to request a rate match or exit the commitment penalty-free. This is aggressive but achievable for large accounts.

5. Batch-first provisions: Negotiate that async/batch workload pricing (50% discount) is guaranteed not to increase for the term. Batch pricing is structurally more stable than on-demand because it allows providers to load-balance compute efficiently.

9. Synthesis: The Strategic Posture

The empirical record is unambiguous: committing to AI inference pricing beyond 3 months has consistently been economically inferior to on-demand purchasing, because the rate of deflation has exceeded the discount rate for all commitment lengths above 1 month.

The correct enterprise posture is:

  1. Pay as you go for frontier models. The deflation rate makes any commitment on GPT-4o-class or Claude Sonnet-class models a bet that prices will not fall, which has been a losing bet in every quarter since 2023.

  2. Use batch APIs for all async workloads. 50% discount, no commitment, full access to future price drops. This is the correct default for document processing, evaluation, overnight analysis, and any workload tolerant of <24h latency.

  3. Prioritize prompt caching over reserved capacity. For workloads with stable system prompts, caching delivers 90% reduction on cached tokens — far exceeding any reserved capacity discount — with no commitment.

  4. Negotiate price-match and model-substitution clauses before signing any commitment. Enterprise volume warrants these provisions. Without them, a commitment is a one-sided bet against deflation.

  5. Budget on volume with forward-price deflation applied. Static-price AI budgets systematically overestimate short-term spend and underestimate the volume growth that cheap tokens enable. Model the forward price curve explicitly.

  6. Monitor per-task cost, not per-token price. As reasoning models become standard, per-token price becomes a less meaningful metric. Track the fully-loaded cost of completing each distinct AI task (summarization, classification, code review, agent action) as the primary KPI.

  7. Treat open-weight models as the floor check. For any workload consuming >10M tokens/day on budget-tier models, run a quarterly make-vs-buy analysis comparing hosted API cost against self-hosting a Llama 4 or Qwen 3.5 class model on spot cloud GPU capacity. The crossover point is moving in favor of self-hosting for high-volume, low-latency workloads.

Sources