How AI Model Prices Moved in H1 2026 (What We Tracked Changing)
We don't publish AI model pricing as a one-off comparison post and walk away from it. We maintain a live reference file — ai-models.ts — that powers our AI Cost Calculator, and every model entry in it carries a sourceNote field documenting exactly what changed (or didn't) the last time we checked it against official pricing pages and independent benchmarks. That discipline exists so the calculator doesn't quietly go stale. It also means the file itself is a genuine, dated record of how this market moved.
This post is that record, pulled straight from our own changelog notes rather than a fresh round of research. Every figure below is what our file says today, and every "before" figure is what our file said before we caught the change. No estimates, no rounding for effect — where our own notes flag a number as unresolved or provisional, we say so instead of picking whichever one sounds cleaner.
- Claude Sonnet 5 launched 2026-06-30 as Anthropic's new default model on Claude Free and Pro plans, at $2/$10 per 1M tokens — an introductory rate that rises to $3/$15 after 2026-08-31, which is exactly what the model it replaced (Sonnet 4.6) already cost. The "cheaper" story has an expiration date.
- Three figures in our own file were simply wrong until this pass corrected them: Grok Code Fast 1's price ($1/$2 listed vs. $0.20/$1.50 actual — a 5x overstatement on input), Llama 4 Scout's price and context (a wrong $0.18/$0.59 and 1M vs. the actual $0.08/$0.30 and a genuinely enormous 10M-token window), and Codestral 25.08's context window (256K listed vs. a real 32K — the correction ran the other direction, revealing less capability, not more).
- Gemini 3.1 Pro has a second pricing tier our calculator doesn't yet surface: prompts over 200K tokens are billed at double the rate ($4/$18 per 1M vs. $2/$12 for shorter prompts).
- DeepSeek is planning a confirmed, real, but not-yet-live repricing for mid-July 2026 — peak/off-peak API pricing that doubles the rate during Beijing business hours. Separately, its V4 Pro tier's 75%-off promotional pricing already expired 2026-05-31, so "DeepSeek is basically free" is already out of date for the Pro tier.
- This is a snapshot of what one actively-maintained internal file caught, not a complete market history — see the caveats section before you treat it as exhaustive.
$2 / $10
Claude Sonnet 5's introductory price per 1M tokens (in/out) — rises to $3/$15 after 2026-08-31
5x
How much Grok Code Fast 1's input price was overstated in our file before this pass
10M
Llama 4 Scout's real context window — the old figure said 1M, missing its entire headline feature
2x
Gemini 3.1 Pro's price multiplier once a prompt crosses 200K tokens — a tier our calculator doesn't show yet
Where This Data Actually Comes From
ai-models.ts is the single source of truth behind our AI Cost Calculator — every model's input/output price, context window, and benchmark scores live there in one typed file. The file carries a top-level AI_MODELS_VERIFIED_DATE stamp (currently 2026-07-06) and, per model, a sourceNote field whose entire job is to say what changed since the previous check: a price correction, a new model added, a model marked legacy, or simply "confirmed unchanged." We're not writing this post from a fresh web search — we're reading our own changelog back and reporting exactly what it says, which is a different (and in some ways more honest) exercise than a normal comparison roundup.
Source: src/lib/data/ai-models.ts, 22 models across Anthropic, OpenAI, Google, DeepSeek, xAI, Mistral, Perplexity, and Meta (via Together AI). The file has been re-verified on at least three dated passes we can point to directly from its own notes: 2026-06-29, 2026-07-04, and 2026-07-06 (the current pass). Every change below is described in the sourceNote of the specific model it applies to — we didn't infer or reconstruct anything from outside the file.
- "Corrected" figures = the model's
sourceNoteexplicitly states a prior value was wrong and names the new, confirmed value. We report both numbers so the size of the correction is visible, not just the end state. - "New/superseded" model status = read directly from the
legacyboolean field plus the model's ownsourceNoteexplaining why. - "Upcoming, not yet live" items = the file explicitly flags these as confirmed real by multiple independent sources but not reflected in current pricing — we kept that distinction rather than treating an announced future price as if it were already in effect.
- Ambiguous or unresolved figures (for example, one model's SWE-bench Verified score being reported as two different numbers by different sources) are left blank in the underlying file rather than guessed at, and we've carried that same "don't guess" rule into this post — see the caveats section near the end.
Why this reads more like "the last six weeks" than all of H1
The Timeline: Every Dated Change Our File Caught
In date order, as documented in the file's own sourceNote fields:
H1 2026 → early H2, as caught by our own reference file
The pre-correction baseline
Claude Sonnet 4.6 ($3/$15 per 1M) was Anthropic's standard mid-flagship. DeepSeek V4 Pro carried a promotional rate roughly 75% below its eventual standard price. Neither fact is independently re-dated in our file beyond "before it changed" — we're not asserting an exact start date for either.
DeepSeek V4 Pro's promo pricing expires
The 75%-off introductory rate ends; pricing reverts to the post-promo rate of $0.435 input / $0.87 output per 1M tokens, still confirmed as current in this pass.
Claude Opus 4.8 re-verified, unchanged
Confirmed still #1 overall and #1 on coding on LMArena's human-preference leaderboard (~1510 Elo), at its existing $5/$25 price. This date is the file's own prior checkpoint for Opus 4.8 specifically.
Claude Sonnet 5 launches, becomes the new default
Confirmed directly on Anthropic's own announcement as the default model for Claude Free and Pro plans, priced at $2/$10 per 1M tokens introductory through 2026-08-31 (rising to $3/$15 after). Claude Sonnet 4.6 is superseded and marked legacy the same day.
Prior data snapshot
The baseline our 2026-07-06 pass diffed against — the point several of the corrections below were measured from.
A cluster of corrections and one addition
Grok Code Fast 1's price corrected ($1/$2 → $0.20/$1.50). Llama 4 Scout's price and context both corrected ($0.18/$0.59 → $0.08/$0.30; 1M → 10M tokens). Codestral 25.08's context corrected (256K → 32K). Gemini 3.1 Pro's second pricing tier discovered (>200K tokens billed at $4/$18). GPT-5.4 mini got its first real, sourced benchmark numbers instead of a placeholder. Claude Sonnet 5 was formally added to the file — six days after its actual launch.
DeepSeek's planned peak/off-peak repricing
Multiple independent sources (including TechNode) confirm DeepSeek plans an "official" V4 launch introducing peak/off-peak API pricing — 2x current rates during Beijing business hours (roughly 9am–12pm and 2pm–6pm local), with off-peak rates unchanged. As of this pass, it hasn't shipped.
Three Numbers That Were Simply Wrong Until This Pass
The most useful thing an actively-maintained reference file can catch isn't a genuine market change — it's its own mistake. Three of the corrections in this pass aren't the market moving at all; they're our file (and, before this pass, plenty of other summaries floating around with the same wrong numbers) being wrong about a model that hadn't changed.
Corrections Caught in the 2026-07-06 Pass
| Model | Field | Previously Listed | Actual, Confirmed | Direction of Error |
|---|---|---|---|---|
| Grok Code Fast 1 | Price (in/out per 1M) | $1.00 / $2.00 | $0.20 / $1.50 | Overstated — input was listed at 5x the real price |
| Llama 4 Scout | Price (in/out per 1M) | $0.18 / $0.59 | $0.08 / $0.30 | Overstated — both figures were roughly 2x the real price |
| Llama 4 Scout | Context window | 1M tokens | 10M tokens (confirmed by two independent sources) | Understated — the model's actual headline feature was invisible in the old data |
| Codestral 25.08 | Context window | 256K tokens | 32K tokens | Overstated — the real window is 1/8th what was listed |
The Codestral correction runs the opposite direction from the other two
Claude's New Default: Sonnet 5 Replaces Sonnet 4.6
The single biggest lineup change in this window: Claude Sonnet 5 launched 2026-06-30 and immediately became the default model on Claude's Free and Pro plans, with Sonnet 4.6 marked legacy the same day. Our own note on the launch is careful to flag that Sonnet 5's benchmark case against Opus 4.8 is genuinely mixed, not a clean win — its SWE-bench Verified score is reported as either 82.1% or 85.2% depending on source (we left it blank in the underlying file rather than pick one), and it trails Opus 4.8 by about 6 points on SWE-bench Pro (63.2% vs. 69.2%), while reportedly beating Opus 4.8 on Terminal-Bench 2.1 and a knowledge-work benchmark called GDPval-AA v2. The fair characterization, in our own note's words: "close to Opus 4.8, meaningfully cheaper (40-60% per multiple sources), not uniformly better."
Sonnet 4.6 vs. Sonnet 5, Same Provider Tier
| Claude Sonnet 4.6 (now legacy) | Claude Sonnet 5 (new default) | |
|---|---|---|
| Price per 1M tokens (in/out) | $3 / $15 | $2 / $10 introductory (through 2026-08-31) |
| Price after 2026-08-31 | $3 / $15 (unchanged) | $3 / $15 — the same number Sonnet 4.6 already cost |
| Context window | 1M tokens | 1M tokens |
| Default status | Superseded 2026-06-30 | Default on Claude Free and Pro plans |
| SWE-bench Pro | Not independently found this pass | 63.2% (vs. Opus 4.8's 69.2%) |
The "cheaper" price has an expiration date
Two Pricing Structures That Aren't a Flat Rate
Two separate findings in this pass point at the same underlying trend: "$X per 1M tokens" is increasingly not the whole story for a given model. It can depend on how long your prompt is, or what time of day you call the API.
Gemini 3.1 Pro's Two Pricing Tiers
| Prompt Length | Input / Output per 1M Tokens |
|---|---|
| ≤200K tokens | $2 / $12 |
| >200K tokens | $4 / $18 — double |
Our calculator's model list doesn't currently surface this second tier — it's confirmed directly against Google's own pricing page this pass, but if your use case regularly sends prompts past 200K tokens, the real cost is double what a flat "$2/$12" comparison implies. Separately, we chased down a genuine source-confusion issue while verifying this model: some sources describe Gemini 3.1 Pro as having a 2M-token context window. Google's own DeepMind model card for this specific model states 1M, unchanged — the 2M figure appears to actually describe two different models, Gemini 3.1 Ultra (a separate model released April 2026) and Gemini 3.5 Pro (whose headline feature is a 2M window, announced at Google I/O in May 2026). We kept Gemini 3.1 Pro at 1M pending clearer primary-source confirmation specifically for that model — a reminder that "context window" claims get conflated across a provider's own model family almost as often as prices do.
DeepSeek's confirmed, not-yet-live mid-July repricing works the same way in a different dimension: the same model, same token count, priced differently depending on the hour you call it. Once live, DeepSeek V4 Flash and V4 Pro will reportedly cost double during Beijing business hours (~9am–12pm and 2pm–6pm local) versus off-peak — with off-peak staying at today's rate. That's on top of V4 Pro's promotional pricing already having expired 2026-05-31 (it reverted from a 75%-off rate to the current $0.435/$0.87 standard rate). "DeepSeek is the cheap option" is still true today, but it's about to become a more conditional statement than it currently is.
One more thing our file caught: a benchmark upgrade, not a price change
What This Means If You're Picking a Model for Client Work
Don't trust last quarter's "cheapest model" list, including ours before this pass
Check context-window claims with the same skepticism as price
A new "default" model can arrive within days and still lag behind in your own tools
Watch for tiered and time-of-day pricing, not just a headline rate
Introductory pricing is a real category — check the expiration date
What This Snapshot Doesn't Cover
- This is not an exhaustive market history of AI pricing in H1 2026. It's what one actively-maintained internal file — the same one powering our AI Cost Calculator — caught changing, documented in its own dated notes.
- The file's dated evidence trail concentrates in a roughly six-week window (2026-05-31 through 2026-07-06), because that's the cadence our re-verification passes actually cover. It doesn't carry equivalent dated proof for January through April 2026, even though some facts referenced here (like Sonnet 4.6's original pricing) were already true earlier in the year.
- We only track a fixed roster of models relevant to the calculator — Anthropic, OpenAI, Google, DeepSeek, xAI, Mistral, Perplexity, and Meta/Together. Real H1 2026 pricing activity from providers outside that roster isn't represented here at all.
- Some figures in our own file are explicitly left unresolved rather than guessed at: Sonnet 5's exact SWE-bench Verified score (82.1% vs. 85.2%, conflicting sources), and Claude Fable 5's 95% SWE-bench Verified figure, sourced from a single site (morphllm.com) and not cross-checked against a second. We're surfacing those as open questions, not settled numbers.
- GPQA Diamond is close to saturated at the frontier right now — Gemini 3.1 Pro's 94.3% and Claude Opus 4.8's 93.6% are close enough to be a statistical tie, not a meaningful gap, per our own benchmark notes.
Related Reading
FAQ / Citation Info
Frequently Asked Questions
- Claude Sonnet 5 launched 2026-06-30 at an introductory $2/$10 per 1M tokens, replacing Sonnet 4.6 ($3/$15) as Anthropic's default Free/Pro model — the intro price rises to $3/$15 after 2026-08-31
- Grok Code Fast 1's real price ($0.20/$1.50 per 1M) was overstated by roughly 5x on input tokens in data until this pass caught it
- Llama 4 Scout's real context window is 10M tokens — 10x what was previously listed, and its actual headline feature
- Codestral 25.08's real context window is 32K tokens — 1/8th of what was previously listed, the correction running the opposite direction from Scout's
- Gemini 3.1 Pro bills prompts over 200K tokens at double the rate ($4/$18 vs. $2/$12 per 1M) — a tier easy to miss on a headline-rate comparison
- DeepSeek's confirmed mid-July 2026 repricing will double API costs during Beijing business hours — not live yet as of this pass