Prompt Engineering Statistics 2026: 80+ Sourced Numbers

Datadog’s telemetry from more than a thousand of its LLM Observability customers found that 69% of all input tokens in customer traces were system prompts: internal instructions, policy definitions and tool guidance. The user’s actual question is the minority of what goes into the window. That is the shape of the skill in production: most of the instruction is written once, by the company, and re-sent on every call.

I pulled together 80 prompt engineering statistics for 2026 and took each one back to the organization that produced it, or to the named report that first published it where the original posting has since come down. Grand View Research, Mordor Intelligence and Fortune Business Insights for market sizing. Lightcast and Stanford HAI for hiring. McKinsey, Deloitte, the World Economic Forum and Microsoft for enterprise adoption. Datadog for production telemetry. Wei et al., Schulhoff et al. and Wharton for the research on techniques themselves.

A word on why this page exists in the shape it does. A page ranking in the top five for this exact query states that North America’s share of the prompt engineering market exceeded $133 billion in 2024, and then, further down that same page, values the entire global market at $0.85 billion in 2024 (SQ Magazine, both figures read on 9 September 2026). A regional slice cannot be more than 150 times the global whole it belongs to. Every figure below is traced to the organization that publishes it, and the Sources and Methodology section at the end lists what was left out and why.

Key Takeaways

• Most of the prompt is written by the company, not the person asking. 69% of input tokens across instrumented production LLM calls are system prompts rather than user input, and among models that support prompt caching only 28% of calls read anything from the cache.
• The market is small and the estimates disagree by more than tenfold. Grand View puts 2026 at $893.7 million and Fortune Business Insights at $673.6 million, while Mordor’s broader category, which folds in agent programming tools, reads $6.95 billion for 2025, about fourteen times Fortune’s figure for that same year.
• Hiring for the exact phrase grew 350%. US AI job postings citing prompt engineering went from 1,393 in 2023 to 6,263 in 2024.
• Employers moved on to agentic AI. The agentic AI share of US postings grew more than 280% year over year, from 0.06% to 0.23%, which works out to roughly 90,000 postings in 2025, while ChatGPT, Conversational AI and Chatbot clusters all declined.
• AI adoption is near-universal, workflow redesign is not. McKinsey’s November 2025 wave puts regular AI use in at least one business function at 88% of organizations, while its earlier wave found only 21% of respondents using gen AI said their organizations had fundamentally redesigned at least some workflows.
• The same prompt does not give the same answer. Wharton found that swapping “Please” for “I order” moved accuracy on individual questions by up to 60 percentage points in either direction, though the differences balanced out across the full dataset.

Top Prompt Engineering Statistics for 2026

If you only quote five numbers from this page, quote these.

Each is stated again in context further down, with its full source note.

• 69% of all input tokens in instrumented production traces were system prompts rather than user input (Datadog, State of AI Engineering 2026, Fact 4).
• 6,263 US AI job postings cited prompt engineering in 2024, up from 1,393 in 2023 (Lightcast data in the Stanford HAI AI Index Report 2025, Figure 4.2.5).
• $893.7 million is Grand View Research’s projection for the global prompt engineering market in 2026 (Grand View Research).
• 88% of organizations report regular AI use in at least one business function (McKinsey, The State of AI, November 2025 wave).
• 63% of organizations sending Datadog LLM telemetry use OpenAI models, down from a 75% share a year ago (Datadog, Fact 1).

Prompt Engineering Market Size Statistics

At least four named analyst houses publish a prompt engineering market figure, and on 2025 they range from about half a billion dollars to seven billion. Mordor’s number is about fourteen times Fortune Business Insights’ for that same year. The gap is almost entirely about where each firm draws the boundary rather than about the underlying activity.

What The Analyst Houses Actually Say

1. Grand View Research valued the global prompt engineering market at $222.1 million in 2023 (Grand View Research).

2. The same firm projects $893.7 million for 2026 (Grand View Research).

3. Grand View projects $2,060.8 million by 2030 (Grand View Research).

4. Grand View states a 32.8% compound annual growth rate for the 2024 to 2030 window (Grand View Research). Its published figures do not sit on a single curve: $893.7 million in 2026 compounding to $2,060.8 million by 2030 works out at 23.2% a year, and the 2023 base compounding to that same 2030 figure works out at 37.5%. The rate and the endpoints come from different vintages of one model, so quote one or the other, never both as though they agree.

5. Fortune Business Insights sizes the same market at $505.43 million for 2025 (Fortune Business Insights).

6. Fortune Business Insights puts 2026 at $673.6 million (Fortune Business Insights).

7. Fortune Business Insights forecasts $6,703.84 million by 2034 (Fortune Business Insights).

8. That forecast carries a 33.27% CAGR across 2026 to 2034 (Fortune Business Insights).

9. Mordor Intelligence sizes a deliberately wider category, prompt engineering and agent programming tools together, at $6.95 billion in 2025 (Mordor Intelligence).

10. Mordor forecasts that category at $40.87 billion by 2030 (Mordor Intelligence).

11. Mordor’s implied growth rate is 42.52% CAGR (Mordor Intelligence).

12. The Business Research Company sizes the market at $1.13 billion in 2025 growing to $1.49 billion in 2026, a step it labels a 32.3% CAGR though the rounded endpoints imply 31.9%, and $4.51 billion by 2030 at a 31.9% CAGR thereafter (The Business Research Company).

The honest way to use these is as a range with the scope attached, not as a single point. Grand View and Fortune Business Insights are measuring prompt engineering tooling. Mordor is measuring that plus LangChain, agent SDKs and the rest of the orchestration layer, which is why its 2025 figure is about fourteen times the Fortune Business Insights figure for the same year. The Business Research Company sits between them. If you need one anchor for a deck, quote a range for 2026 of roughly $0.7 billion to $1.5 billion and say out loud that every figure in it is an analyst model, not revenue anyone reported.

Where The Money Sits Inside The Market

13. Software took more than 71% of component share in 2023 (Grand View Research).

14. North America held 34.0% of revenue share in 2023 (Grand View Research).

15. The United States accounts for approximately 38% of the global market (Fortune Business Insights).

16. The US market specifically went from $61.4 million in 2023 to a projected $546.1 million by 2030 (Grand View Research). Grand View labels this a 32.1% CAGR for 2024 to 2030, but those two endpoints imply 36.6% a year, so the stated rate runs off a 2024 base the page does not publish. Use the dollar figures rather than the growth rate.

17. Generated knowledge prompting holds roughly 42% of technique share (Fortune Business Insights).

Prompt Engineering Jobs And Salary Statistics

Job postings are the cleanest signal in this category, because Lightcast indexes online postings at scale and publishes its skill taxonomy. Read the denominator carefully: the counts in the next subsection are drawn from US AI job postings, not from all US job postings, so they describe competition inside AI hiring rather than the whole labor market. The figures in the subsection after that switch to the all-postings denominator, and say so. Postings are also not hires, and employers tag skills inconsistently, so the direction is more trustworthy than any single count.

AI Postings That Name The Skill

18. US AI job postings citing prompt engineering: 1,393 in 2023, 6,263 in 2024, an increase of 350% (Stanford HAI AI Index Report 2025, Lightcast data, Figure 4.2.5).

19. US AI job postings citing generative AI: 15,741 in 2023, 66,635 in 2024, up 323% (Stanford HAI AI Index Report 2025, Figure 4.2.5).

20. US AI job postings citing large language modeling: 4,956 in 2023, 19,562 in 2024, up 295% (Stanford HAI AI Index Report 2025, Figure 4.2.5).

21. Prompt engineering appeared in 5.68% of US AI job postings in 2024, up from 4.62% in 2023 (Stanford HAI AI Index Report 2025, Figure 4.2.6).

Where AI Hiring Went Next

22. Agentic AI skill mentions jumped from 0.06% of all US job postings in 2024 to 0.23% in 2025 (Lightcast).

23. That 0.23% share covers roughly 90,000 US job postings, a share increase of more than 280% year over year (Lightcast).

24. AI skills appeared in 2.5% of all US job postings in 2025, a 55% increase from the prior year (Lightcast).

25. ChatGPT, Conversational AI and Chatbot skill clusters all declined from 2024 to 2025 (Lightcast).

That last one is the finding worth sitting with. Demand for AI skills is climbing hard, and the specific vocabulary of the chat era is falling out of postings. The skill has not died. It has been folded into roles that assume you can write a working system prompt, evaluate the output and iterate, without anyone writing “prompt engineer” on the requisition.

What A Prompt Engineer Actually Got Paid

There is no audited industry median for this role, and anyone quoting one is quoting a crowdsourced number without a disclosed sample. One first-party disclosure holds up.

26. Anthropic’s Prompt Engineer and Librarian listing offered a base salary range of $175,000 to $335,000 in San Francisco (Fortune, March 9, 2023).

Treat Anthropic’s disclosed range as one employer’s number for one unusual role in 2023, not as a market rate. Two higher figures circulate alongside it; both are handled in Sources and Methodology.

AI Adoption Statistics

Who Is Actually Using AI

27. 88% of organizations report regular AI use in at least one business function, up from 78% a year earlier (McKinsey, The State of AI, November 2025 wave, n=1,993 across 105 nations, fielded June 25 to July 29, 2025).

28. In the prior edition, 71% of respondents said their organizations regularly use gen AI in at least one business function, up from 65% in early 2024 (McKinsey, March 2025).

29. 78% of organizations reported using AI in 2024, up from 55% in 2023 (Stanford HAI AI Index Report 2025, Chapter 4, which republishes McKinsey survey data rather than measuring adoption itself).

30. Four in five university students now use generative AI (Stanford HAI AI Index Report 2026).

Who Inside The Company Is Actually Prompting

31. 53% of C-level executives said they regularly use gen AI at work, against 44% of midlevel managers (McKinsey, March 2025 edition, fielded July 2024).

32. Organizations most often used gen AI in marketing and sales, ahead of product and service development, service operations and software engineering (McKinsey, March 2025 edition, fielded July 2024).

33. 21% of respondents reporting gen AI use say their organizations have fundamentally redesigned at least some workflows (McKinsey, March 2025).

34. In the November 2025 wave that split into 20% of ordinary respondents against 55% of AI high performers (McKinsey).

Marketing and sales being the most common deployment function is the reason this page belongs on a content marketing site rather than an engineering one. The people writing the most consequential prompts in most companies are not engineers. I wrote up how that actually plays out day to day in prompt engineering for content marketing.

Money And Friction

35. US private AI investment reached $109.1 billion in 2024 (Stanford HAI AI Index Report 2025, Chapter 4).

36. Private investment in generative AI specifically reached $33.9 billion in 2024, up 18.7% from 2023 (Stanford HAI AI Index Report 2025).

37. 78% of respondents expect to increase overall AI spending in the next fiscal year (Deloitte, State of Generative AI in the Enterprise Q4 2024, n=2,773 director to C-suite leaders across 14 countries).

38. 69% say fully implementing a governance strategy will take over a year (Deloitte).

39. Regulatory compliance rose from 28% to 38% as the top barrier between the first and fourth survey waves of 2024 (Deloitte).

40. Workforce access to sanctioned AI tools grew from fewer than 40% to about 60% in a year (Deloitte, State of AI in the Enterprise 2026, n=3,235 across 24 countries).

41. Only 21% of companies have mature agent governance models (Deloitte 2026).

42. 25% report AI having a transformative effect, double the prior year (Deloitte 2026).

Both McKinsey and Deloitte sell AI advisory services, so read their optimism with that in mind. The friction numbers are the more useful half of both surveys, because a consultancy has less incentive to invent a governance backlog than to invent enthusiasm.

Prompting Technique Statistics From The Research

Chain Of Thought And What It Actually Did

43. PaLM 540B scored 56.9% on GSM8K math word problems with chain-of-thought prompting against 17.9% with standard prompting, a gain of 39.0 points (Wei et al., Google Research, NeurIPS 2022, Table 2).

44. The effect only appears at scale, roughly 100 billion parameters and above (Wei et al., Figure 4).

Cite this one as history, not as advice. Modern reasoning models train chain-of-thought behavior internally, which means the technique that produced a 39-point jump in 2022 is largely baked into the model you are using now.

How Many Techniques There Actually Are

45. A systematic survey catalogued 58 distinct LLM prompting techniques (Schulhoff et al., The Prompt Report).

46. The same survey documented 40 techniques for other modalities (Schulhoff et al.).

47. It also standardized 33 vocabulary terms for a field that had been using words inconsistently (Schulhoff et al.).

How Much The Same Prompt Varies

48. Wharton’s Generative AI Labs ran 100 repetitions per condition against GPT-4o and GPT-4o-mini on the GPQA Diamond benchmark (Wharton GAIL, March 4, 2025).

49. Saying “Please” instead of “I order” shifted performance on individual questions by up to 60 percentage points in either direction, though the differences balanced out across the full dataset (Wharton GAIL).

That balancing-out clause matters and gets dropped every time this study is quoted. Politeness is not a magic lever. What the study actually shows is that any single prompt-and-question pair carries enormous variance, which is the empirical case for testing your prompts against a set of real inputs instead of eyeballing one good answer.

Context Engineering Statistics

In June 2025 the vocabulary shifted. Tobi Lütke, Shopify’s CEO, wrote on 18 June 2025 that he liked “the term ‘context engineering’ over prompt engineering” because “it describes the core skill better: the art of providing all the context for the task to be plausibly solvable by the LLM.” A week later Andrej Karpathy endorsed it on 25 June 2025, writing that “in every industrial-strength LLM app, context engineering is the delicate art and science of filling the context window with just the right information for the next step.”

Datadog’s production telemetry is the evidence that the rename described something real.

What Is Actually In The Context Window

50. 69% of all input tokens in customer traces were system prompts: internal instructions, policy definitions and tool guidance (Datadog, Fact 4).

51. Among models that support prompt caching, only 28% of LLM call spans show any cached-read input tokens (Datadog, Fact 4).

52. The average number of tokens per request more than doubled year over year for median customers (Datadog, Fact 5).

53. Ninetieth-percentile power users quadrupled their tokens per request over the same period (Datadog, Fact 5).

Statistics 50 and 51 together are the most actionable pair on this page. Roughly two thirds of what you send is scaffolding rather than the user’s question, and even among models that support caching, nearly three quarters of calls read nothing from the cache. Datadog measures message roles and cache reads, not whether the text is byte-identical between calls, so treat this as a strong signal to check your own caching rather than proof about any one application.

The Multi-Model Reality

54. 63% of organizations sending Datadog LLM telemetry use OpenAI models, down from a 75% share of organizations a year ago. Datadog notes this is not a fall in absolute use: the number of its customers using OpenAI more than doubled over the same period (Datadog, Fact 1).

55. Google Gemini and Anthropic Claude gained 20 and 23 percentage points respectively over the last year (Datadog, Fact 1).

56. More than 70% of organizations now use three or more models (Datadog, Fact 1).

57. The share of organizations using more than six models nearly doubled (Datadog, Fact 1).

A prompt that works on one model is not a prompt that works, which is worth remembering when you pick tools from the AI content marketing stack. If most organizations are running three or more models, the portable artifact is the context you assemble, not the phrasing you tuned against one provider.

How Agentic Systems Actually Behave

58. Agent framework adoption nearly doubled year over year, rising from more than 9% of organizations in early 2025 to almost 18% by the beginning of 2026 (Datadog, Fact 3).

59. 59% of agentic application requests made only a single service call (Datadog, Fact 7).

60. Only 18% of end-to-end agentic application requests made three or more service calls (Datadog, Fact 7).

Where LLM Calls Fail

61. In February 2026, 5% of all LLM call spans reported an error and 60% of those errors were caused by exceeded rate limits (Datadog, Fact 6).

62. In March 2026, 2% of spans returned an error, with rate limits accounting for almost a third of them, nearly 8.4 million rate-limit errors in total (Datadog, Fact 6).

Datadog sells observability tooling, and the report argues for its own product category. The methodology disclosure is still the most detailed in this corpus: per-fact method notes, a stated character-count proxy for tokens, and a CC BY-ND license. The sample skews toward companies serious enough about production AI to instrument it, which means these numbers describe competent teams rather than the average one.

AI Skills And Training Statistics

What Employers Say They Need

63. 86% of employers expect AI and information processing technologies to transform their business by 2030 (World Economic Forum, Future of Jobs Report 2025, 1,000+ employers representing over 14 million workers across 55 economies).

64. AI and big data top the list of fastest-growing skills for 2025 to 2030, followed by networks and cybersecurity, then technological literacy (World Economic Forum).

65. 63% of surveyed employers name skills gaps as the top barrier to business transformation (World Economic Forum).

66. 82% of leaders say this is a pivotal year to rethink key aspects of strategy and operations (Microsoft, 2025 Work Trend Index, n=31,000 across 31 markets, administered by Edelman Data x Intelligence).

What People Are Actually Learning

67. Over 3.2 million of nearly 7.4 million AI enrollments on Coursera in 2024 were in generative AI training (Coursera).

68. That works out to an average of six enrollments per minute in 2024, against two per minute in 2023, with India and the United States leading (Coursera).

69. Generative AI enrollments have since reached 14 per minute, the most in-demand skill in Coursera’s history (Coursera Job Skills Report 2026).

70. Enterprise-learner generative AI enrollments grew 234% year over year (Coursera Job Skills Report 2026).

Coursera is reporting on its own platform, so this is a measure of Coursera demand rather than of world demand. It is still the only enrollment series in this category with a consistent method across three years, and the shape of it, a sevenfold increase in pace between 2023 and 2026, matches what the job posting data shows.

What The Work Itself Looks Like

71. The average worker receives 117 emails daily and 153 Teams messages per weekday (Microsoft, Breaking Down the Infinite Workday, June 2025).

72. 58% of AI users say they are producing work they could not have produced a year ago, rising to 80% among the most AI-fluent group (Microsoft, 2026 Work Trend Index, n=20,000 across 10 countries).

73. Only 19% of AI users fall into Microsoft’s “Frontier” tier, where organizational capability and individual readiness are both high (Microsoft, 2026 Work Trend Index).

Microsoft sells the productivity stack it is measuring, and the telemetry comes from Microsoft 365 customers only. I read these as directional.

What The Model Vendors Report

Model providers publish two kinds of numbers: benchmark results for their own prompting patterns, and customer outcomes their customers agreed to have quoted. Both are vendor-run and neither is independently audited. The benchmark numbers at least name the benchmark, which is more than the customer stories do.

Prompt Patterns In Vendor Documentation

74. OpenAI reports that its own agentic scaffold for GPT-4.1 solves 55% of problems on SWE-bench Verified, and recommends three types of reminder in all agent prompts (OpenAI, GPT-4.1 Prompting Guide, April 2025).

75. GPT-5 introduced a “minimal” reasoning effort tier, the fastest option that still uses the reasoning paradigm (OpenAI, GPT-5 Prompting Guide, August 2025).

76. Anthropic announced Agent Skills on October 16, 2025 and published them as an open standard for cross-platform portability on December 18, 2025 (Anthropic).

Both vendor guides say the same uncomfortable thing in different words: prompts need re-engineering when models change generation. Anything you write today is a depreciating asset.

Vendor-Reported Customer Outcomes

77. Stripe has enabled Claude Code across 1,370 engineers (Anthropic).

78. One Stripe team migrated 10,000 lines of Scala to Java in four days against an estimated ten engineering weeks by hand (Anthropic).

79. Rakuten cut average time to market for new features from 24 working days to 5, a 79% reduction (Anthropic).

80. Ramp reports early observations of up to 80% faster initial incident triage (Anthropic).

These are vendor-published customer outcomes with no independent audit and no disclosed measurement window. They are worth quoting as what the vendor claims, and they are not evidence of what you will get.

Frequently Asked Questions

How Many Job Postings Mention Prompt Engineering?

US AI job postings citing “prompt engineering” grew from 1,393 in 2023 to 6,263 in 2024, 350% growth, according to Lightcast data published in the Stanford HAI AI Index Report 2025. The denominator matters: those are AI job postings, not all US postings, and prompt engineering appeared in 5.68% of them. For context, AI postings naming generative AI went from 15,741 to 66,635 over the same period. The widely repeated “1,200% year over year” figure does not match any measured dataset and should not be used.

What Is The Prompt Engineering Market Worth In 2026?

It depends entirely on scope. Grand View Research projects $893.7 million for 2026 and Fortune Business Insights projects $673.6 million, both measuring prompt engineering tooling. The Business Research Company projects $1.49 billion. Mordor Intelligence reads $6.95 billion for 2025 because its category also includes agent programming tools like LangChain and agent SDKs. The “$1.52 billion in 2026” figure that circulates online matches none of these published series. The nearest named-analyst figure for 2026 is The Business Research Company’s $1.49 billion, so cite that firm and its exact number instead.

What Is The Average Prompt Engineer Salary?

There is no audited industry median, and the crowdsourced figures in circulation do not disclose their sample sizes. The cleanest first-party disclosure is Anthropic’s Prompt Engineer and Librarian listing at $175,000 to $335,000 base in San Francisco, confirmed by Fortune’s March 2023 reporting. Use it as one employer’s number for one unusual role, not as a market rate. The often-quoted $525,000 figure belongs to a different, unnamed employer.

What Share Of LLM Input Tokens Are System Prompts?

69%, according to Datadog’s State of AI Engineering 2026, which analyzed telemetry from more than a thousand of its LLM Observability customers. Those tokens are internal instructions, policy definitions and tool guidance rather than the user’s own question. Among models that support prompt caching, only 28% of LLM call spans show any cached-read tokens. Datadog measures message roles and cache reads on call spans, not whether any given application re-sends identical text, so read this as a prompt to check your own caching rather than a claim about applications in general.

Did Chain-Of-Thought Prompting Actually Improve Accuracy?

Yes, at scale and on reasoning benchmarks. Wei et al. showed PaLM 540B reaching 56.9% on GSM8K with chain-of-thought prompting against 17.9% with standard prompting. The effect only emerged at roughly 100 billion parameters and above. Modern reasoning models internalize this behavior during training, so the technique that once produced a 39-point jump is now largely built in.

Is Prompt Engineering Still A Real Job In 2026?

The page has no measurement of how many roles carry the title itself, so treat what follows as a description of where the skill sits rather than a verdict on the job market. Lightcast’s 2026 data shows ChatGPT, Conversational AI and Chatbot skill clusters all declining from 2024 to 2025 while the agentic AI share of postings grew more than 280% year over year, from 0.06% to 0.23%, which works out to roughly 90,000 postings. Pure prompt-writing roles have been absorbed into broader AI engineering and applied roles that assume you can write a system prompt, evaluate output and iterate.

Why Do Prompt Engineering Statistics Vary So Much Between Sources?

Three reasons. Scope, when analysts include or exclude agent tooling, which produces the tenfold market-size gap. Recency, when a 2024 survey wave gets quoted as though it describes 2026. And provenance, when an aggregator repeats a figure that has no underlying study. The last category is the most common, and the fastest test is to click the citation and see whether it leads to an organization that ran a survey or to another blog post. The same problem runs through the content marketing statistics and SEO statistics circuits.

Sources And Methodology

Every figure on this page was taken back to the organization that produced it in September 2026, and in almost every case the number was read off that organization’s own page or PDF, or, where the original posting has come down or the page is gated, off the named report that first published it. Two things are worth stating plainly. First, McKinsey’s site refused direct automated requests, so its figures were read from McKinsey’s own hosted report file rather than from the landing page the citation links to; no independent second source confirms them, because the adoption figures the Stanford HAI AI Index publishes are McKinsey survey data republished, not a separate measurement. Second, Grand View Research publishes a 2023 base, a 2026 estimate and a 2030 projection that do not sit on one growth curve; that is noted inline rather than smoothed over. Where a live fetch returned the number, the citation links to the page that carried it.

• Datadog, “State of AI Engineering 2026,” 2026 (the report page carries no explicit publication date; its latest data is March 2026). Aggregated, anonymized telemetry from more than a thousand LLM Observability customers, data through March 2026, per-fact methodology disclosed, character count used as a token proxy, CC BY-ND 4.0. https://www.datadoghq.com/state-of-ai-engineering/
• Stanford HAI, “Artificial Intelligence Index Report 2025,” April 2025, and the 2026 edition. Compilation of primary datasets including Lightcast. Job-posting counts come from the Chapter 4 PDF, Figures 4.2.5 and 4.2.6, not the landing page. https://hai.stanford.edu/assets/files/hai_ai-index-report-2025_chapter4_final.pdf ; https://hai.stanford.edu/ai-index/2025-ai-index-report ; https://hai.stanford.edu/ai-index/2026-ai-index-report
• Lightcast, Stanford AI Index 2025 and 2026 contributions. Full universe of indexed US online job postings, approximately 300 AI skills tracked. https://lightcast.io/resources/research/stanford-ai-index-2026
• McKinsey and Company (QuantumBlack), “The State of AI,” November 2025 wave, n=1,993 respondents across 105 nations, fielded June 25 to July 29, 2025, and the March 2025 edition. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
• Deloitte, “State of Generative AI in the Enterprise Q4 2024,” January 2025, n=2,773 across 14 countries, and “State of AI in the Enterprise 2026,” January 2026, n=3,235 across 24 countries fielded August to September 2025. https://www.deloitte.com/us/en/about/press-room/state-of-generative-ai.html ; https://www.deloitte.com/us/en/about/press-room/state-of-ai-report-2026.html
• World Economic Forum, “Future of Jobs Report 2025,” January 2025. 1,000+ employers representing over 14 million workers across 55 economies. https://reports.weforum.org/docs/WEF_Future_of_Jobs_Report_2025.pdf
• Coursera, contribution to the WEF Future of Jobs Report 2025 and the Job Skills Report 2026. Platform enrollment data. https://blog.coursera.org/wef-future-of-jobs-report-2025/
• Microsoft, “2025 Annual Work Trend Index,” April 23, 2025, n=31,000 across 31 markets administered by Edelman Data x Intelligence; “Breaking Down the Infinite Workday,” June 2025; and the 2026 Work Trend Index, n=20,000 across 10 countries. https://www.microsoft.com/en-us/worklab/work-trend-index/2025-the-year-the-frontier-firm-is-born ; https://www.microsoft.com/en-us/worklab/work-trend-index/breaking-down-infinite-workday ; https://www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization
• Wei, Wang, Schuurmans, et al., “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models,” Google Research, NeurIPS 2022. https://arxiv.org/abs/2201.11903
• Schulhoff, Ilie, Balepur, et al., “The Prompt Report: A Systematic Survey of Prompting Techniques,” current revision. https://arxiv.org/abs/2406.06608
• Wharton Generative AI Labs, “Prompt Engineering is Complicated and Contingent,” March 4, 2025. GPT-4o and GPT-4o-mini on GPQA Diamond, 100 repetitions per condition. https://gail.wharton.upenn.edu/research-and-insights/tech-report-prompt-engineering-is-complicated-and-contingent/
• Grand View Research, “Prompt Engineering Market Size and Share Report, 2030.” Analyst-built market model, methodology gated. https://www.grandviewresearch.com/industry-analysis/prompt-engineering-market-report
• Mordor Intelligence, “Prompt Engineering and Agent Programming Tools Market.” Broader category including agent programming tools. https://www.mordorintelligence.com/industry-reports/prompt-engineering-and-agent-programming-tools-market
• Fortune Business Insights, “Prompt Engineering Market 2026 to 2034.” https://www.fortunebusinessinsights.com/prompt-engineering-market-109382
• The Business Research Company, “Prompt Engineering Global Market Report,” 2026 edition. https://www.thebusinessresearchcompany.com/report/prompt-engineering-global-market-report
• Fortune, “The hottest new job in tech,” March 9, 2023. The only live source for Anthropic’s original disclosed salary range; the listing itself has been taken down. https://fortune.com/2023/03/09/new-ai-jobs-chatgpt-like-assistants/
• Tobi Lütke and Andrej Karpathy on context engineering, June 2025. https://x.com/tobi/status/1935533422589399127 ; https://x.com/karpathy/status/1937902205765607626
• Anthropic customer case studies and engineering blog. Vendor-attested, not independently audited. https://claude.com/customers/stripe ; https://claude.com/customers/rakuten ; https://claude.com/customers/ramp ; https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills
• OpenAI Cookbook, GPT-4.1 and GPT-5 prompting guides. Vendor-run benchmarks. https://cookbook.openai.com/examples/gpt4-1_prompting_guide ; https://cookbook.openai.com/examples/gpt-5/gpt-5_prompting_guide

Figures excluded and why. The following claims appear widely in search results and were left off this page because no organization, survey instrument, sample size or publication could be located for them: “85% of organizations using generative AI say prompt engineering is critical to success”; “prompt engineering job postings increased 1,200% year over year,” which is superseded by Lightcast’s measured 350% increase from 2023 to 2024; “structured prompts reduce errors by up to 76%”; “70% of AI projects fail without prompt engineering,” which appears to be a rephrasing of Gartner cancellation forecasts that do not mention prompt engineering; “62% of AI professionals spend 20% or more of their time on prompt optimization”; “47% of developers include prompt engineering in their core skillset,” which matches neither the Stack Overflow Developer Survey nor GitHub Octoverse.

Three further figures were dropped during verification rather than at the research stage. The “$525,000 Anthropic listing” is a misattribution: that upper bound belongs to a different, unnamed employer. The “1,200% year over year” posting-growth figure has no measured source and is superseded by Lightcast’s measured 350% increase from 2023 to 2024. An updated Anthropic listing at $250,000 to $375,000 base is repeated across many secondary pages, but the listing has been taken down and no live primary source carries it. A frequently cited “PRISMA-style review of 1,565 papers” framing for The Prompt Report does not appear in the paper’s current abstract, so only the taxonomy counts that do appear are used here. Figures reported as “100+ emails and 150+ messages per day” were replaced with the precise 117 and 153 from the Microsoft report that actually published them.

Related figures on adjacent topics are collected in website statistics.

Where an organization has released a newer edition since this article was first researched, the newer number is used and the older one is kept alongside it with its original date, so a reader tracking a figure over time can see both.